Compare commits
3 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
9d05f77a9b | ||
|
|
23e093fc75 | ||
|
|
f77ce4dff2 |
@@ -1,546 +1,197 @@
|
||||
"""Session compaction with structured fact extraction.
|
||||
"""Session compaction with fact extraction.
|
||||
|
||||
Before compressing conversation context, extract durable facts with enough
|
||||
structure to survive retrieval: source/provenance, temporal anchors,
|
||||
normalized canonical keys, and contradiction groups.
|
||||
Before compressing conversation context, extracts durable facts
|
||||
(user preferences, corrections, project details) and saves them
|
||||
to the fact store so they survive compression.
|
||||
|
||||
Usage:
|
||||
from agent.session_compactor import extract_and_save_facts
|
||||
facts = extract_and_save_facts(messages)
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import logging
|
||||
import re
|
||||
import time
|
||||
from dataclasses import dataclass, field
|
||||
from datetime import datetime, timezone
|
||||
from typing import Any, Dict, List, Tuple
|
||||
from typing import Any, Dict, List, Optional, Tuple
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
_DEPLOY_METHOD_RE = re.compile(r"\bdeploy(?:ing)?\s+(?:via|through|with)\s+([A-Za-z0-9_./+-]+)", re.IGNORECASE)
|
||||
_WATCHDOG_CAP_RE = re.compile(
|
||||
r"\b(?:the\s+)?([A-Za-z0-9_-]+(?:\s+watchdog)?)\s+(?:caps|limits)\s+dispatches(?:\s+per\s+cycle)?\s+to\s+([0-9]+)",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
_PROVIDER_RE = re.compile(
|
||||
r"\bprovider\s+(?:is|should\s+stay|should\s+be|needs\s+to\s+be)\s+([A-Za-z0-9._/-]+)",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
_MODEL_RE = re.compile(
|
||||
r"\bmodel\s+(?:is|should\s+stay|should\s+be|needs\s+to\s+be)\s+([A-Za-z0-9._:/-]+)",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
_PORT_RE = re.compile(r"\bport\s+(?:is|should\s+be)\s+([0-9]+)", re.IGNORECASE)
|
||||
_PROJECT_USES_RE = re.compile(r"\b(?:the\s+)?project\s+(?:uses|needs|requires)\s+(.+?)(?:[.!?]|$)", re.IGNORECASE)
|
||||
_PREFERENCE_RE = re.compile(r"\bI\s+(?:prefer|like|want|need)\s+(.+?)(?:[.!?]|$)", re.IGNORECASE)
|
||||
_CONSTRAINT_RE = re.compile(r"\b(?:do\s+not|don't)\s+(?:ever\s+|again\s+)?(.+?)(?:[.!?]|$)", re.IGNORECASE)
|
||||
_DECISION_RE = re.compile(r"\b(?:we|the\s+team)\s+(?:decided|agreed|chose)\s+(?:to\s+)?(.+?)(?:[.!?]|$)", re.IGNORECASE)
|
||||
|
||||
|
||||
@dataclass
|
||||
class ExtractedFact:
|
||||
"""A durable fact extracted from conversation."""
|
||||
|
||||
category: str
|
||||
entity: str
|
||||
content: str
|
||||
confidence: float
|
||||
source_turn: int
|
||||
"""A fact extracted from conversation."""
|
||||
category: str # "user_pref", "correction", "project", "tool_quirk", "general"
|
||||
entity: str # what the fact is about
|
||||
content: str # the fact itself
|
||||
confidence: float # 0.0-1.0
|
||||
source_turn: int # which message turn it came from
|
||||
timestamp: float = 0.0
|
||||
source_role: str = "user"
|
||||
source_text: str = ""
|
||||
normalized_content: str = ""
|
||||
canonical_key: str = ""
|
||||
relation: str = "general"
|
||||
contradiction_group: str = ""
|
||||
status: str = "active"
|
||||
provenance: str = ""
|
||||
observed_at: str = ""
|
||||
evidence: List[Dict[str, Any]] = field(default_factory=list)
|
||||
metadata: Dict[str, Any] = field(default_factory=dict)
|
||||
|
||||
def __post_init__(self) -> None:
|
||||
if not self.timestamp:
|
||||
self.timestamp = time.time()
|
||||
if not self.observed_at:
|
||||
self.observed_at = _iso_from_timestamp(self.timestamp)
|
||||
if not self.normalized_content:
|
||||
self.normalized_content = _normalize_value(self.content)
|
||||
if not self.provenance:
|
||||
self.provenance = f"conversation:{self.source_role}:{self.source_turn}"
|
||||
if not self.canonical_key:
|
||||
self.canonical_key = _canonical_key(self.entity, self.relation, self.normalized_content)
|
||||
if not self.evidence:
|
||||
self.evidence = [
|
||||
{
|
||||
"source_role": self.source_role,
|
||||
"source_turn": self.source_turn,
|
||||
"source_text": self.source_text or self.content,
|
||||
"observed_at": self.observed_at,
|
||||
"provenance": self.provenance,
|
||||
}
|
||||
]
|
||||
self.metadata = dict(self.metadata or {})
|
||||
self.metadata.setdefault("entity", self.entity)
|
||||
self.metadata.setdefault("relation", self.relation)
|
||||
self.metadata.setdefault("value", self.content)
|
||||
self.metadata.setdefault("normalized_value", self.normalized_content)
|
||||
self.metadata.setdefault("provenance", [self.provenance])
|
||||
self.metadata.setdefault("evidence", list(self.evidence))
|
||||
self.metadata.setdefault("observation_count", len(self.evidence))
|
||||
self.metadata.setdefault("duplicate_count", max(0, self.metadata["observation_count"] - 1))
|
||||
if self.contradiction_group:
|
||||
self.metadata.setdefault("contradiction_group", self.contradiction_group)
|
||||
self.metadata.setdefault("status", self.status)
|
||||
|
||||
# Patterns that indicate user preferences
|
||||
_PREFERENCE_PATTERNS = [
|
||||
(r"(?:I|we) (?:prefer|like|want|need) (.+?)(?:\.|$)", "preference"),
|
||||
(r"(?:always|never) (?:use|do|run|deploy) (.+?)(?:\.|$)", "preference"),
|
||||
(r"(?:my|our) (?:default|preferred|usual) (.+?) (?:is|are) (.+?)(?:\.|$)", "preference"),
|
||||
(r"(?:make sure|ensure|remember) (?:to|that) (.+?)(?:\.|$)", "instruction"),
|
||||
(r"(?:don'?t|do not) (?:ever|ever again) (.+?)(?:\.|$)", "constraint"),
|
||||
]
|
||||
|
||||
# Patterns that indicate corrections
|
||||
_CORRECTION_PATTERNS = [
|
||||
(r"(?:actually|no[, ]|wait[, ]|correction[: ]|sorry[, ]) (.+)", "correction"),
|
||||
(r"(?:I meant|what I meant was|the correct) (.+?)(?:\.|$)", "correction"),
|
||||
(r"(?:it'?s|its) (?:not|shouldn'?t be|wrong) (.+?)(?:\.|$)", "correction"),
|
||||
]
|
||||
|
||||
# Patterns that indicate project/tool facts
|
||||
_PROJECT_PATTERNS = [
|
||||
(r"(?:the |our )?(?:project|repo|codebase|code) (?:is|uses|needs|requires) (.+?)(?:\.|$)", "project"),
|
||||
(r"(?:deploy|push|commit) (?:to|on) (.+?)(?:\.|$)", "project"),
|
||||
(r"(?:this|that|the) (?:server|host|machine|VPS) (?:is|runs|has) (.+?)(?:\.|$)", "infrastructure"),
|
||||
(r"(?:model|provider|engine) (?:is|should be|needs to be) (.+?)(?:\.|$)", "config"),
|
||||
]
|
||||
|
||||
|
||||
def extract_facts_from_messages(messages: List[Dict[str, Any]]) -> List[ExtractedFact]:
|
||||
"""Extract durable facts from conversation messages.
|
||||
|
||||
Scans conversation turns for preferences, decisions, corrections, and
|
||||
operational state. Raw candidates are normalized into canonical facts so
|
||||
near-duplicates merge and contradictions remain inspectable.
|
||||
Scans user messages for preferences, corrections, project facts,
|
||||
and infrastructure details that should survive compression.
|
||||
"""
|
||||
facts = []
|
||||
seen_contents = set()
|
||||
|
||||
raw_candidates: list[ExtractedFact] = []
|
||||
for turn_idx, msg in enumerate(messages):
|
||||
role = msg.get("role", "")
|
||||
content = msg.get("content", "")
|
||||
if role not in {"user", "assistant"}:
|
||||
|
||||
# Only scan user messages and assistant responses with corrections
|
||||
if role not in ("user", "assistant"):
|
||||
continue
|
||||
if not content or not isinstance(content, str):
|
||||
continue
|
||||
if len(content) < 10:
|
||||
continue
|
||||
|
||||
# Skip tool results and system messages
|
||||
if role == "assistant" and msg.get("tool_calls"):
|
||||
continue
|
||||
if not isinstance(content, str) or len(content.strip()) < 10:
|
||||
continue
|
||||
|
||||
timestamp, observed_at = _message_time(msg)
|
||||
raw_candidates.extend(
|
||||
_extract_from_text(
|
||||
content.strip(),
|
||||
turn_idx=turn_idx,
|
||||
role=role,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
)
|
||||
)
|
||||
extracted = _extract_from_text(content, turn_idx, role)
|
||||
|
||||
return _normalize_candidates(raw_candidates)
|
||||
|
||||
|
||||
def evaluate_extraction_quality(messages: List[Dict[str, Any]]) -> Dict[str, Any]:
|
||||
"""Return before/after metrics for raw vs normalized extraction quality."""
|
||||
|
||||
raw_candidates: list[ExtractedFact] = []
|
||||
for turn_idx, msg in enumerate(messages):
|
||||
role = msg.get("role", "")
|
||||
content = msg.get("content", "")
|
||||
if role not in {"user", "assistant"}:
|
||||
continue
|
||||
if role == "assistant" and msg.get("tool_calls"):
|
||||
continue
|
||||
if not isinstance(content, str) or len(content.strip()) < 10:
|
||||
continue
|
||||
timestamp, observed_at = _message_time(msg)
|
||||
raw_candidates.extend(
|
||||
_extract_from_text(
|
||||
content.strip(),
|
||||
turn_idx=turn_idx,
|
||||
role=role,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
)
|
||||
)
|
||||
|
||||
normalized = _normalize_candidates(raw_candidates)
|
||||
raw_count = len(raw_candidates)
|
||||
normalized_count = len(normalized)
|
||||
contradiction_groups = {
|
||||
fact.contradiction_group
|
||||
for fact in normalized
|
||||
if fact.status == "contradiction" and fact.contradiction_group
|
||||
}
|
||||
duplicate_count = max(0, raw_count - normalized_count)
|
||||
noise_reduction = (duplicate_count / raw_count) if raw_count else 0.0
|
||||
|
||||
return {
|
||||
"raw_candidates": raw_count,
|
||||
"normalized_facts": normalized_count,
|
||||
"duplicates_merged": duplicate_count,
|
||||
"contradiction_groups": len(contradiction_groups),
|
||||
"noise_reduction": round(noise_reduction, 3),
|
||||
}
|
||||
|
||||
|
||||
def _extract_from_text(
|
||||
text: str,
|
||||
*,
|
||||
turn_idx: int,
|
||||
role: str,
|
||||
timestamp: float,
|
||||
observed_at: str,
|
||||
) -> List[ExtractedFact]:
|
||||
"""Extract raw fact candidates from a single text block."""
|
||||
|
||||
facts: list[ExtractedFact] = []
|
||||
if role != "user":
|
||||
return facts
|
||||
|
||||
deploy_match = _DEPLOY_METHOD_RE.search(text)
|
||||
if deploy_match:
|
||||
method = deploy_match.group(1).strip()
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="project.decision",
|
||||
entity="project",
|
||||
relation="workflow.deploy_method",
|
||||
value=method,
|
||||
content=f"Deploy via {method}",
|
||||
confidence=0.88,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=True,
|
||||
)
|
||||
)
|
||||
|
||||
watchdog_match = _WATCHDOG_CAP_RE.search(text)
|
||||
if watchdog_match:
|
||||
watchdog = watchdog_match.group(1).strip()
|
||||
cap = watchdog_match.group(2).strip()
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="project.operational",
|
||||
entity=_normalize_entity(watchdog),
|
||||
relation="fleet.dispatch_cap",
|
||||
value=cap,
|
||||
content=f"{watchdog} caps dispatches per cycle to {cap}",
|
||||
confidence=0.92,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=True,
|
||||
)
|
||||
)
|
||||
|
||||
provider_match = _PROVIDER_RE.search(text)
|
||||
if provider_match:
|
||||
provider = provider_match.group(1).strip()
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="project.config",
|
||||
entity="project",
|
||||
relation="config.provider",
|
||||
value=provider,
|
||||
content=f"Provider should stay {provider}",
|
||||
confidence=0.91,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=True,
|
||||
)
|
||||
)
|
||||
|
||||
model_match = _MODEL_RE.search(text)
|
||||
if model_match:
|
||||
model = model_match.group(1).strip()
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="project.config",
|
||||
entity="project",
|
||||
relation="config.model",
|
||||
value=model,
|
||||
content=f"Model should stay {model}",
|
||||
confidence=0.9,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=True,
|
||||
)
|
||||
)
|
||||
|
||||
port_match = _PORT_RE.search(text)
|
||||
if port_match:
|
||||
port = port_match.group(1).strip()
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="project.config",
|
||||
entity="project",
|
||||
relation="config.port",
|
||||
value=port,
|
||||
content=f"Port is {port}",
|
||||
confidence=0.9,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=True,
|
||||
)
|
||||
)
|
||||
|
||||
project_match = _PROJECT_USES_RE.search(text)
|
||||
if project_match:
|
||||
value = project_match.group(1).strip().rstrip(".")
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="project.stack",
|
||||
entity="project",
|
||||
relation="project.stack",
|
||||
value=value,
|
||||
content=f"Project uses {value}",
|
||||
confidence=0.74,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=False,
|
||||
)
|
||||
)
|
||||
|
||||
preference_match = _PREFERENCE_RE.search(text)
|
||||
if preference_match:
|
||||
value = preference_match.group(1).strip().rstrip(".")
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="user_pref.preference",
|
||||
entity="user",
|
||||
relation="user.preference",
|
||||
value=value,
|
||||
content=value,
|
||||
confidence=0.72,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=False,
|
||||
)
|
||||
)
|
||||
|
||||
constraint_match = _CONSTRAINT_RE.search(text)
|
||||
if constraint_match:
|
||||
value = constraint_match.group(1).strip().rstrip(".")
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="user_pref.constraint",
|
||||
entity="user",
|
||||
relation="user.constraint",
|
||||
value=value,
|
||||
content=f"Do not {value}",
|
||||
confidence=0.82,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=False,
|
||||
)
|
||||
)
|
||||
|
||||
decision_match = _DECISION_RE.search(text)
|
||||
if decision_match:
|
||||
value = decision_match.group(1).strip().rstrip(".")
|
||||
facts.append(
|
||||
_build_fact(
|
||||
category="project.decision",
|
||||
entity="project",
|
||||
relation="project.decision",
|
||||
value=value,
|
||||
content=f"Decision: {value}",
|
||||
confidence=0.79,
|
||||
source_turn=turn_idx,
|
||||
source_role=role,
|
||||
source_text=text,
|
||||
timestamp=timestamp,
|
||||
observed_at=observed_at,
|
||||
unique_slot=False,
|
||||
)
|
||||
)
|
||||
# Deduplicate by content
|
||||
for fact in extracted:
|
||||
key = f"{fact.category}:{fact.content[:100]}"
|
||||
if key not in seen_contents:
|
||||
seen_contents.add(key)
|
||||
facts.append(fact)
|
||||
|
||||
return facts
|
||||
|
||||
|
||||
def _build_fact(
|
||||
*,
|
||||
category: str,
|
||||
entity: str,
|
||||
relation: str,
|
||||
value: str,
|
||||
content: str,
|
||||
confidence: float,
|
||||
source_turn: int,
|
||||
source_role: str,
|
||||
source_text: str,
|
||||
timestamp: float,
|
||||
observed_at: str,
|
||||
unique_slot: bool,
|
||||
) -> ExtractedFact:
|
||||
normalized_value = _normalize_value(value.rstrip(".!?"))
|
||||
value = value.rstrip(".!?")
|
||||
content = content.rstrip(".!?")
|
||||
provenance = f"conversation:{source_role}:{source_turn}"
|
||||
contradiction_group = relation if unique_slot else ""
|
||||
evidence = [
|
||||
{
|
||||
"source_role": source_role,
|
||||
"source_turn": source_turn,
|
||||
"source_text": source_text,
|
||||
"observed_at": observed_at,
|
||||
"provenance": provenance,
|
||||
}
|
||||
]
|
||||
metadata = {
|
||||
"entity": entity,
|
||||
"relation": relation,
|
||||
"value": value,
|
||||
"normalized_value": normalized_value,
|
||||
"provenance": [provenance],
|
||||
"evidence": list(evidence),
|
||||
"observation_count": 1,
|
||||
"duplicate_count": 0,
|
||||
"status": "active",
|
||||
}
|
||||
if contradiction_group:
|
||||
metadata["contradiction_group"] = contradiction_group
|
||||
return ExtractedFact(
|
||||
category=category,
|
||||
entity=entity,
|
||||
content=content,
|
||||
confidence=confidence,
|
||||
source_turn=source_turn,
|
||||
timestamp=timestamp,
|
||||
source_role=source_role,
|
||||
source_text=source_text,
|
||||
normalized_content=normalized_value,
|
||||
canonical_key=_canonical_key(entity, relation, normalized_value),
|
||||
relation=relation,
|
||||
contradiction_group=contradiction_group,
|
||||
status="active",
|
||||
provenance=provenance,
|
||||
observed_at=observed_at,
|
||||
evidence=evidence,
|
||||
metadata=metadata,
|
||||
)
|
||||
def _extract_from_text(text: str, turn_idx: int, role: str) -> List[ExtractedFact]:
|
||||
"""Extract facts from a single text block."""
|
||||
facts = []
|
||||
timestamp = time.time()
|
||||
|
||||
# Clean text for pattern matching
|
||||
clean = text.strip()
|
||||
|
||||
def _normalize_candidates(candidates: List[ExtractedFact]) -> List[ExtractedFact]:
|
||||
"""Merge duplicates and mark contradictions while preserving evidence."""
|
||||
# User preference patterns (from user messages)
|
||||
if role == "user":
|
||||
for pattern, subcategory in _PREFERENCE_PATTERNS:
|
||||
for match in re.finditer(pattern, clean, re.IGNORECASE):
|
||||
content = match.group(1).strip() if match.lastindex else match.group(0).strip()
|
||||
if len(content) > 5:
|
||||
facts.append(ExtractedFact(
|
||||
category=f"user_pref.{subcategory}",
|
||||
entity="user",
|
||||
content=content[:200],
|
||||
confidence=0.7,
|
||||
source_turn=turn_idx,
|
||||
timestamp=timestamp,
|
||||
))
|
||||
|
||||
by_key: dict[str, ExtractedFact] = {}
|
||||
contradiction_groups: dict[str, list[ExtractedFact]] = {}
|
||||
# Correction patterns (from user messages)
|
||||
if role == "user":
|
||||
for pattern, subcategory in _CORRECTION_PATTERNS:
|
||||
for match in re.finditer(pattern, clean, re.IGNORECASE):
|
||||
content = match.group(1).strip() if match.lastindex else match.group(0).strip()
|
||||
if len(content) > 5:
|
||||
facts.append(ExtractedFact(
|
||||
category=f"correction.{subcategory}",
|
||||
entity="user",
|
||||
content=content[:200],
|
||||
confidence=0.8,
|
||||
source_turn=turn_idx,
|
||||
timestamp=timestamp,
|
||||
))
|
||||
|
||||
for candidate in candidates:
|
||||
existing = by_key.get(candidate.canonical_key)
|
||||
if existing is not None:
|
||||
by_key[candidate.canonical_key] = _merge_fact(existing, candidate)
|
||||
continue
|
||||
# Project/infrastructure patterns (from both user and assistant)
|
||||
for pattern, subcategory in _PROJECT_PATTERNS:
|
||||
for match in re.finditer(pattern, clean, re.IGNORECASE):
|
||||
content = match.group(1).strip() if match.lastindex else match.group(0).strip()
|
||||
if len(content) > 5:
|
||||
facts.append(ExtractedFact(
|
||||
category=f"project.{subcategory}",
|
||||
entity=subcategory,
|
||||
content=content[:200],
|
||||
confidence=0.6,
|
||||
source_turn=turn_idx,
|
||||
timestamp=timestamp,
|
||||
))
|
||||
|
||||
by_key[candidate.canonical_key] = candidate
|
||||
if candidate.contradiction_group:
|
||||
contradiction_groups.setdefault(candidate.contradiction_group, []).append(candidate)
|
||||
|
||||
for group, facts in contradiction_groups.items():
|
||||
canonical_keys = {fact.canonical_key for fact in facts}
|
||||
if len(canonical_keys) <= 1:
|
||||
continue
|
||||
for fact in facts:
|
||||
fact.status = "contradiction"
|
||||
fact.metadata["status"] = "contradiction"
|
||||
fact.metadata["contradiction_group"] = group
|
||||
fact.metadata["contradiction_keys"] = sorted(canonical_keys - {fact.canonical_key})
|
||||
|
||||
return sorted(by_key.values(), key=lambda fact: (fact.source_turn, fact.timestamp, fact.canonical_key))
|
||||
|
||||
|
||||
def _merge_fact(existing: ExtractedFact, incoming: ExtractedFact) -> ExtractedFact:
|
||||
existing.confidence = max(existing.confidence, incoming.confidence)
|
||||
existing.timestamp = min(existing.timestamp, incoming.timestamp)
|
||||
existing.source_turn = min(existing.source_turn, incoming.source_turn)
|
||||
if not existing.observed_at or (incoming.observed_at and incoming.observed_at < existing.observed_at):
|
||||
existing.observed_at = incoming.observed_at
|
||||
existing.provenance = min(existing.provenance, incoming.provenance)
|
||||
|
||||
provenance = _ordered_unique(existing.metadata.get("provenance", []), incoming.metadata.get("provenance", []))
|
||||
evidence = _merge_evidence(existing.metadata.get("evidence", []), incoming.metadata.get("evidence", []))
|
||||
observation_count = int(existing.metadata.get("observation_count", len(existing.evidence) or 1))
|
||||
observation_count += int(incoming.metadata.get("observation_count", len(incoming.evidence) or 1))
|
||||
|
||||
existing.evidence = evidence
|
||||
existing.metadata["provenance"] = provenance
|
||||
existing.metadata["evidence"] = evidence
|
||||
existing.metadata["observation_count"] = observation_count
|
||||
existing.metadata["duplicate_count"] = max(0, observation_count - 1)
|
||||
existing.metadata["status"] = existing.status
|
||||
return existing
|
||||
return facts
|
||||
|
||||
|
||||
def save_facts_to_store(facts: List[ExtractedFact], fact_store_fn=None) -> int:
|
||||
"""Save extracted facts to the fact store.
|
||||
|
||||
If a callback is supplied, prefer the structured signature but fall back to
|
||||
the legacy four-argument callback for compatibility.
|
||||
Args:
|
||||
facts: List of extracted facts.
|
||||
fact_store_fn: Optional callable(category, entity, content, trust).
|
||||
If None, uses the holographic fact store if available.
|
||||
|
||||
Returns:
|
||||
Number of facts saved.
|
||||
"""
|
||||
|
||||
saved = 0
|
||||
for fact in facts:
|
||||
payload = {
|
||||
"category": _store_category(fact.category),
|
||||
"entity": fact.entity,
|
||||
"content": fact.content,
|
||||
"trust": fact.confidence,
|
||||
"metadata": dict(fact.metadata),
|
||||
"canonical_key": fact.canonical_key,
|
||||
"observed_at": fact.observed_at,
|
||||
"source_role": fact.source_role,
|
||||
"source_turn": fact.source_turn,
|
||||
"contradiction_group": fact.contradiction_group,
|
||||
"status": fact.status,
|
||||
"relation": fact.relation,
|
||||
}
|
||||
|
||||
if fact_store_fn:
|
||||
if fact_store_fn:
|
||||
for fact in facts:
|
||||
try:
|
||||
fact_store_fn(**payload)
|
||||
fact_store_fn(
|
||||
category=fact.category,
|
||||
entity=fact.entity,
|
||||
content=fact.content,
|
||||
trust=fact.confidence,
|
||||
)
|
||||
saved += 1
|
||||
continue
|
||||
except TypeError:
|
||||
try:
|
||||
fact_store_fn(payload["category"], payload["entity"], payload["content"], payload["trust"])
|
||||
saved += 1
|
||||
continue
|
||||
except Exception as exc:
|
||||
logger.debug("Failed to save fact via callback: %s", exc)
|
||||
continue
|
||||
except Exception as exc:
|
||||
logger.debug("Failed to save fact via callback: %s", exc)
|
||||
continue
|
||||
|
||||
except Exception as e:
|
||||
logger.debug("Failed to save fact: %s", e)
|
||||
else:
|
||||
# Try holographic fact store
|
||||
try:
|
||||
from fact_store import fact_store as _fs
|
||||
|
||||
tags = ",".join(filter(None, [fact.entity, fact.relation, fact.status]))
|
||||
_fs(
|
||||
action="add",
|
||||
content=fact.content,
|
||||
category=_store_category(fact.category),
|
||||
tags=tags,
|
||||
trust_delta=fact.confidence - 0.5,
|
||||
)
|
||||
saved += 1
|
||||
for fact in facts:
|
||||
try:
|
||||
_fs(
|
||||
action="add",
|
||||
content=fact.content,
|
||||
category=fact.category,
|
||||
tags=fact.entity,
|
||||
trust_delta=fact.confidence - 0.5,
|
||||
)
|
||||
saved += 1
|
||||
except Exception as e:
|
||||
logger.debug("Failed to save fact via fact_store: %s", e)
|
||||
except ImportError:
|
||||
logger.debug("fact_store not available — facts not persisted")
|
||||
break
|
||||
except Exception as exc:
|
||||
logger.debug("Failed to save fact via fact_store: %s", exc)
|
||||
|
||||
return saved
|
||||
|
||||
@@ -553,10 +204,9 @@ def extract_and_save_facts(
|
||||
|
||||
Returns (extracted_facts, saved_count).
|
||||
"""
|
||||
|
||||
facts = extract_facts_from_messages(messages)
|
||||
if facts:
|
||||
logger.info("Extracted %d normalized facts from conversation", len(facts))
|
||||
logger.info("Extracted %d facts from conversation", len(facts))
|
||||
saved = save_facts_to_store(facts, fact_store_fn)
|
||||
logger.info("Saved %d/%d facts to store", saved, len(facts))
|
||||
else:
|
||||
@@ -566,105 +216,16 @@ def extract_and_save_facts(
|
||||
|
||||
def format_facts_summary(facts: List[ExtractedFact]) -> str:
|
||||
"""Format extracted facts as a readable summary."""
|
||||
|
||||
if not facts:
|
||||
return "No facts extracted."
|
||||
|
||||
by_category: dict[str, list[ExtractedFact]] = {}
|
||||
for fact in facts:
|
||||
by_category.setdefault(fact.category, []).append(fact)
|
||||
by_category = {}
|
||||
for f in facts:
|
||||
by_category.setdefault(f.category, []).append(f)
|
||||
|
||||
lines = [f"Extracted {len(facts)} facts:", ""]
|
||||
for category, category_facts in sorted(by_category.items()):
|
||||
lines.append(f" {category}:")
|
||||
for fact in category_facts:
|
||||
suffix = f" [{fact.status}]" if fact.status != "active" else ""
|
||||
lines.append(f" - {fact.content[:80]}{suffix}")
|
||||
for cat, cat_facts in sorted(by_category.items()):
|
||||
lines.append(f" {cat}:")
|
||||
for f in cat_facts:
|
||||
lines.append(f" - {f.content[:80]}")
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
def _store_category(category: str) -> str:
|
||||
if category.startswith("user_pref"):
|
||||
return "user_pref"
|
||||
if category.startswith("project"):
|
||||
return "project"
|
||||
if category.startswith("tool"):
|
||||
return "tool"
|
||||
return "general"
|
||||
|
||||
|
||||
def _message_time(msg: Dict[str, Any]) -> Tuple[float, str]:
|
||||
for key in ("created_at", "timestamp", "time"):
|
||||
value = msg.get(key)
|
||||
if value is None:
|
||||
continue
|
||||
if isinstance(value, (int, float)):
|
||||
ts = float(value)
|
||||
return ts, _iso_from_timestamp(ts)
|
||||
if isinstance(value, str):
|
||||
parsed = _parse_time_string(value)
|
||||
if parsed is not None:
|
||||
return parsed, _iso_from_timestamp(parsed) if "T" not in value else value.replace("+00:00", "Z")
|
||||
return time.time(), value
|
||||
now = time.time()
|
||||
return now, _iso_from_timestamp(now)
|
||||
|
||||
|
||||
def _parse_time_string(value: str) -> float | None:
|
||||
text = value.strip()
|
||||
if not text:
|
||||
return None
|
||||
try:
|
||||
return float(text)
|
||||
except ValueError:
|
||||
pass
|
||||
try:
|
||||
normalized = text[:-1] + "+00:00" if text.endswith("Z") else text
|
||||
return datetime.fromisoformat(normalized).timestamp()
|
||||
except ValueError:
|
||||
return None
|
||||
|
||||
|
||||
def _iso_from_timestamp(value: float) -> str:
|
||||
return datetime.fromtimestamp(value, tz=timezone.utc).isoformat().replace("+00:00", "Z")
|
||||
|
||||
|
||||
def _normalize_value(value: str) -> str:
|
||||
normalized = re.sub(r"[^a-z0-9]+", " ", value.lower())
|
||||
normalized = re.sub(r"\s+", " ", normalized).strip()
|
||||
return normalized
|
||||
|
||||
|
||||
def _normalize_entity(value: str) -> str:
|
||||
return _normalize_value(value).replace(" ", "_") or "entity"
|
||||
|
||||
|
||||
def _canonical_key(entity: str, relation: str, normalized_value: str) -> str:
|
||||
return f"{entity}|{relation}|{normalized_value}"
|
||||
|
||||
|
||||
def _ordered_unique(*groups: List[str]) -> List[str]:
|
||||
seen: set[str] = set()
|
||||
ordered: list[str] = []
|
||||
for group in groups:
|
||||
for item in group:
|
||||
if item and item not in seen:
|
||||
seen.add(item)
|
||||
ordered.append(item)
|
||||
return ordered
|
||||
|
||||
|
||||
def _merge_evidence(existing: List[Dict[str, Any]], incoming: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
|
||||
seen: set[tuple[str, str, str]] = set()
|
||||
merged: list[dict[str, Any]] = []
|
||||
for item in list(existing) + list(incoming):
|
||||
key = (
|
||||
str(item.get("provenance", "")),
|
||||
str(item.get("observed_at", "")),
|
||||
str(item.get("source_text", "")),
|
||||
)
|
||||
if key in seen:
|
||||
continue
|
||||
seen.add(key)
|
||||
merged.append(dict(item))
|
||||
return merged
|
||||
|
||||
@@ -1,194 +1,757 @@
|
||||
[
|
||||
{
|
||||
"id": "screenshot_github_home",
|
||||
"id": "screenshot_github_mark",
|
||||
"url": "https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": ["github", "logo", "mark"],
|
||||
"expected_keywords": [
|
||||
"github",
|
||||
"logo",
|
||||
"mark"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "diagram_mermaid_flow",
|
||||
"url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6siSZXVhjQTlgl1nigHg5fRBOzSfebopROCu_cytObSfgLSE1ANOeZWkO2IH5upZxYot8m1hqAdpD_63WRl0xdUG1jdl9kPiOb_EWk2JBtPaiKkF4eVIYgO0EtkW-RSgC4gJ6HJYRG1UNdN0HNVd0Bftjj7X8P92qPj-F8l8T3w",
|
||||
"id": "screenshot_github_social",
|
||||
"url": "https://github.githubassets.com/images/modules/site/social-cards.png",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"github",
|
||||
"page",
|
||||
"web"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_github_code_search",
|
||||
"url": "https://github.githubassets.com/images/modules/site/features-code-search.png",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"search",
|
||||
"code",
|
||||
"feature"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_terminal_capture",
|
||||
"url": "https://raw.githubusercontent.com/nicehash/nicehash-quick-start/main/images/nicehash-terminal.png",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"terminal",
|
||||
"command",
|
||||
"output"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_http_404",
|
||||
"url": "https://http.cat/404.jpg",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"404",
|
||||
"error",
|
||||
"cat"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_dummy_cli_01",
|
||||
"url": "https://dummyimage.com/1280x720/111827/f9fafb.png&text=Hermes+CLI+Session+01",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"hermes",
|
||||
"cli",
|
||||
"session"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_dummy_cli_02",
|
||||
"url": "https://dummyimage.com/1280x720/0f172a/e2e8f0.png&text=Prompt+Cache+Dashboard",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"prompt",
|
||||
"cache",
|
||||
"dashboard"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_dummy_ui_01",
|
||||
"url": "https://dummyimage.com/1280x720/1f2937/f3f4f6.png&text=Settings+Panel+Voice+Mode",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"settings",
|
||||
"voice",
|
||||
"mode"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_dummy_ui_02",
|
||||
"url": "https://dummyimage.com/1280x720/334155/f8fafc.png&text=Browser+Vision+Preview",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"browser",
|
||||
"vision",
|
||||
"preview"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_dummy_ui_03",
|
||||
"url": "https://dummyimage.com/1280x720/111827/ffffff.png&text=Tool+Call+Inspector",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": [
|
||||
"tool",
|
||||
"call",
|
||||
"inspector"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "diagram_flow_a",
|
||||
"url": "https://dummyimage.com/1200x800/f8fafc/0f172a.png&text=Flowchart+API+Gateway+Queue+Worker",
|
||||
"category": "diagram",
|
||||
"expected_keywords": ["flow", "diagram", "process"],
|
||||
"expected_keywords": [
|
||||
"flowchart",
|
||||
"api",
|
||||
"worker"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_1",
|
||||
"url": "https://picsum.photos/seed/vision1/400/300",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"id": "diagram_flow_b",
|
||||
"url": "https://dummyimage.com/1200x800/f1f5f9/0f172a.png&text=Architecture+Diagram+Database+Cache+Client",
|
||||
"category": "diagram",
|
||||
"expected_keywords": [
|
||||
"architecture",
|
||||
"diagram",
|
||||
"cache"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_2",
|
||||
"url": "https://picsum.photos/seed/vision2/400/300",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"id": "diagram_uml_a",
|
||||
"url": "https://dummyimage.com/1200x800/e2e8f0/0f172a.png&text=Class+Diagram+User+Session+Message",
|
||||
"category": "diagram",
|
||||
"expected_keywords": [
|
||||
"class",
|
||||
"diagram",
|
||||
"session"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_simple_bar",
|
||||
"url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['Q1','Q2','Q3','Q4'],datasets:[{label:'Revenue',data:[100,150,200,250]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": ["bar", "chart", "revenue"],
|
||||
"id": "diagram_uml_b",
|
||||
"url": "https://dummyimage.com/1200x800/cbd5e1/0f172a.png&text=Sequence+Diagram+Request+Response",
|
||||
"category": "diagram",
|
||||
"expected_keywords": [
|
||||
"sequence",
|
||||
"diagram",
|
||||
"response"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_pie",
|
||||
"url": "https://quickchart.io/chart?c={type:'pie',data:{labels:['A','B','C'],datasets:[{data:[30,50,20]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": ["pie", "chart", "percentage"],
|
||||
"id": "diagram_network_a",
|
||||
"url": "https://dummyimage.com/1200x800/ffffff/111827.png&text=Network+Nodes+Edges+Router",
|
||||
"category": "diagram",
|
||||
"expected_keywords": [
|
||||
"network",
|
||||
"node",
|
||||
"router"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "diagram_network_b",
|
||||
"url": "https://dummyimage.com/1200x800/ffffff/1e293b.png&text=Service+Mesh+Proxy+Auth",
|
||||
"category": "diagram",
|
||||
"expected_keywords": [
|
||||
"service",
|
||||
"mesh",
|
||||
"auth"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "diagram_state_machine",
|
||||
"url": "https://dummyimage.com/1200x800/f8fafc/334155.png&text=State+Machine+Idle+Run+Stop",
|
||||
"category": "diagram",
|
||||
"expected_keywords": [
|
||||
"state",
|
||||
"machine",
|
||||
"idle"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "diagram_mind_map",
|
||||
"url": "https://dummyimage.com/1200x800/fefce8/1f2937.png&text=Mind+Map+Memory+Recall+Tools",
|
||||
"category": "diagram",
|
||||
"expected_keywords": [
|
||||
"mind",
|
||||
"memory",
|
||||
"tools"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "diagram_pipeline",
|
||||
"url": "https://dummyimage.com/1200x800/ecfeff/155e75.png&text=Pipeline+Ingest+Rank+Summarize",
|
||||
"category": "diagram",
|
||||
"expected_keywords": [
|
||||
"pipeline",
|
||||
"ingest",
|
||||
"summarize"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "diagram_org_chart",
|
||||
"url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
|
||||
"url": "https://dummyimage.com/1200x800/fdf2f8/831843.png&text=Org+Chart+Lead+Review+Ops",
|
||||
"category": "diagram",
|
||||
"expected_keywords": ["organization", "hierarchy", "chart"],
|
||||
"expected_keywords": [
|
||||
"org",
|
||||
"chart",
|
||||
"review"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_terminal",
|
||||
"url": "https://raw.githubusercontent.com/nicehash/nicehash-quick-start/main/images/nicehash-terminal.png",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": ["terminal", "command", "output"],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_3",
|
||||
"url": "https://picsum.photos/seed/vision3/400/300",
|
||||
"id": "photo_random_01",
|
||||
"url": "https://picsum.photos/seed/vision-bench-1/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_line",
|
||||
"id": "photo_random_02",
|
||||
"url": "https://picsum.photos/seed/vision-bench-2/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_03",
|
||||
"url": "https://picsum.photos/seed/vision-bench-3/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_04",
|
||||
"url": "https://picsum.photos/seed/vision-bench-4/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_05",
|
||||
"url": "https://picsum.photos/seed/vision-bench-5/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_06",
|
||||
"url": "https://picsum.photos/seed/vision-bench-6/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_07",
|
||||
"url": "https://picsum.photos/seed/vision-bench-7/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_08",
|
||||
"url": "https://picsum.photos/seed/vision-bench-8/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_09",
|
||||
"url": "https://picsum.photos/seed/vision-bench-9/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_10",
|
||||
"url": "https://picsum.photos/seed/vision-bench-10/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 30,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_bar_quarterly",
|
||||
"url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['Q1','Q2','Q3','Q4'],datasets:[{label:'Revenue',data:[100,150,200,250]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": [
|
||||
"bar",
|
||||
"chart",
|
||||
"revenue"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_pie_market",
|
||||
"url": "https://quickchart.io/chart?c={type:'pie',data:{labels:['A','B','C'],datasets:[{data:[30,50,20]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": [
|
||||
"pie",
|
||||
"chart",
|
||||
"percentage"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_line_temp",
|
||||
"url": "https://quickchart.io/chart?c={type:'line',data:{labels:['Jan','Feb','Mar','Apr'],datasets:[{label:'Temperature',data:[5,8,12,18]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": ["line", "chart", "temperature"],
|
||||
"expected_keywords": [
|
||||
"line",
|
||||
"chart",
|
||||
"temperature"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "diagram_sequence",
|
||||
"url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
|
||||
"category": "diagram",
|
||||
"expected_keywords": ["sequence", "interaction", "message"],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_4",
|
||||
"url": "https://picsum.photos/seed/vision4/400/300",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_webpage",
|
||||
"url": "https://github.githubassets.com/images/modules/site/social-cards.png",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": ["github", "page", "web"],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "chart_radar",
|
||||
"id": "chart_radar_skill",
|
||||
"url": "https://quickchart.io/chart?c={type:'radar',data:{labels:['Speed','Power','Defense','Magic'],datasets:[{label:'Hero',data:[80,60,70,90]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": ["radar", "chart", "skill"],
|
||||
"expected_keywords": [
|
||||
"radar",
|
||||
"chart",
|
||||
"skill"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_5",
|
||||
"url": "https://picsum.photos/seed/vision5/400/300",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "diagram_class",
|
||||
"url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
|
||||
"category": "diagram",
|
||||
"expected_keywords": ["class", "object", "attribute"],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "chart_doughnut",
|
||||
"url": "https://quickchart.io/chart?c={type:'doughnut',data:{labels:['Desktop','Mobile','Tablet'],datasets:[{data:[60,30,10]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": ["doughnut", "chart", "device"],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_6",
|
||||
"url": "https://picsum.photos/seed/vision6/400/300",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_error",
|
||||
"url": "https://http.cat/404.jpg",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": ["404", "error", "cat"],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": true}
|
||||
},
|
||||
{
|
||||
"id": "diagram_network",
|
||||
"url": "https://mermaid.ink/img/pako:eNpdkE9PwzAMxb-K5VOl7gc7sAOIIDuAw9gptnRaSJLSJttQStmXs9LCH-ymBOI1ef_42U6cUSae4IkDxbAAWtB6iuyIWyrLgXLALrPEAfFy-iCcmk-83RSjcFZ-51ac2k7AW0JqAKY9y9IcsAPzdS3jxBb5NrHUAraH_lutjbpi6oJqG7P7IPEd3-ItJsWCaO1FVYLw8qQwANsJbIt8i1AExAX0OCwjNqoa6LoPaq7oCvbHHmv5f7pVfX4K5b8mvg",
|
||||
"category": "diagram",
|
||||
"expected_keywords": ["network", "node", "connection"],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_7",
|
||||
"url": "https://picsum.photos/seed/vision7/400/300",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
},
|
||||
{
|
||||
"id": "chart_stacked_bar",
|
||||
"id": "chart_stacked_cloud",
|
||||
"url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['2022','2023','2024'],datasets:[{label:'Cloud',data:[100,150,200]},{label:'On-prem',data:[200,180,160]}]},options:{scales:{x:{stacked:true},y:{stacked:true}}}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": ["stacked", "bar", "chart"],
|
||||
"expected_keywords": [
|
||||
"stacked",
|
||||
"bar",
|
||||
"chart"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": true}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "screenshot_dashboard",
|
||||
"url": "https://github.githubassets.com/images/modules/site/features-code-search.png",
|
||||
"category": "screenshot",
|
||||
"expected_keywords": ["search", "code", "feature"],
|
||||
"id": "chart_area_growth",
|
||||
"url": "https://quickchart.io/chart?c={type:'line',data:{labels:['W1','W2','W3','W4'],datasets:[{label:'Growth',data:[10,15,18,24],fill:true}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": [
|
||||
"line",
|
||||
"growth",
|
||||
"chart"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "photo_random_8",
|
||||
"url": "https://picsum.photos/seed/vision8/400/300",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"id": "chart_scatter_eval",
|
||||
"url": "https://quickchart.io/chart?c={type:'scatter',data:{datasets:[{label:'Runs',data:[{x:1,y:70},{x:2,y:75},{x:3,y:82}]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": [
|
||||
"scatter",
|
||||
"chart",
|
||||
"runs"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": false}
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_horizontal_bar",
|
||||
"url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['UI','OCR','Docs'],datasets:[{label:'Score',data:[88,76,91]}]},options:{indexAxis:'y'}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": [
|
||||
"bar",
|
||||
"score",
|
||||
"ocr"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_bubble_usage",
|
||||
"url": "https://quickchart.io/chart?c={type:'bubble',data:{datasets:[{label:'Latency',data:[{x:1,y:120,r:8},{x:2,y:95,r:6},{x:3,y:180,r:10}]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": [
|
||||
"bubble",
|
||||
"latency",
|
||||
"chart"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "chart_doughnut_devices",
|
||||
"url": "https://quickchart.io/chart?c={type:'doughnut',data:{labels:['Desktop','Mobile','Tablet'],datasets:[{data:[60,30,10]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": [
|
||||
"doughnut",
|
||||
"chart",
|
||||
"device"
|
||||
],
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {
|
||||
"min_length": 50,
|
||||
"min_sentences": 2,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "ocr_text_01",
|
||||
"url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Hermes+OCR+Alpha+01",
|
||||
"category": "ocr",
|
||||
"expected_keywords": [
|
||||
"hermes",
|
||||
"ocr"
|
||||
],
|
||||
"ground_truth_ocr": "Hermes OCR Alpha 01",
|
||||
"expected_structure": {
|
||||
"min_length": 10,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "ocr_text_02",
|
||||
"url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Prompt+Cache+Hit+87%",
|
||||
"category": "ocr",
|
||||
"expected_keywords": [
|
||||
"prompt",
|
||||
"cache"
|
||||
],
|
||||
"ground_truth_ocr": "Prompt Cache Hit 87%",
|
||||
"expected_structure": {
|
||||
"min_length": 10,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "ocr_text_03",
|
||||
"url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Session+42+Ready",
|
||||
"category": "ocr",
|
||||
"expected_keywords": [
|
||||
"session",
|
||||
"42"
|
||||
],
|
||||
"ground_truth_ocr": "Session 42 Ready",
|
||||
"expected_structure": {
|
||||
"min_length": 10,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "ocr_text_04",
|
||||
"url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Latency+118+ms",
|
||||
"category": "ocr",
|
||||
"expected_keywords": [
|
||||
"latency",
|
||||
"118"
|
||||
],
|
||||
"ground_truth_ocr": "Latency 118 ms",
|
||||
"expected_structure": {
|
||||
"min_length": 10,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "ocr_text_05",
|
||||
"url": "https://dummyimage.com/1200x320/ffffff/000000.png&text=Voice+Mode+Enabled",
|
||||
"category": "ocr",
|
||||
"expected_keywords": [
|
||||
"voice",
|
||||
"mode"
|
||||
],
|
||||
"ground_truth_ocr": "Voice Mode Enabled",
|
||||
"expected_structure": {
|
||||
"min_length": 10,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "document_text_01",
|
||||
"url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Invoice+1001+Total+42+Due+2026-04-22",
|
||||
"category": "document",
|
||||
"expected_keywords": [
|
||||
"invoice",
|
||||
"1001",
|
||||
"total"
|
||||
],
|
||||
"ground_truth_ocr": "Invoice 1001 Total 42 Due 2026-04-22",
|
||||
"expected_structure": {
|
||||
"min_length": 20,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "document_text_02",
|
||||
"url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Form+A+Name+Alice+Status+Approved",
|
||||
"category": "document",
|
||||
"expected_keywords": [
|
||||
"form",
|
||||
"a",
|
||||
"name"
|
||||
],
|
||||
"ground_truth_ocr": "Form A Name Alice Status Approved",
|
||||
"expected_structure": {
|
||||
"min_length": 20,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "document_text_03",
|
||||
"url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Report+Memory+Recall+Score+91+Percent",
|
||||
"category": "document",
|
||||
"expected_keywords": [
|
||||
"report",
|
||||
"memory",
|
||||
"recall"
|
||||
],
|
||||
"ground_truth_ocr": "Report Memory Recall Score 91 Percent",
|
||||
"expected_structure": {
|
||||
"min_length": 20,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "document_text_04",
|
||||
"url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Checklist+Crisis+Escalation+Call+988+Now",
|
||||
"category": "document",
|
||||
"expected_keywords": [
|
||||
"checklist",
|
||||
"crisis",
|
||||
"escalation"
|
||||
],
|
||||
"ground_truth_ocr": "Checklist Crisis Escalation Call 988 Now",
|
||||
"expected_structure": {
|
||||
"min_length": 20,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"id": "document_text_05",
|
||||
"url": "https://dummyimage.com/1400x900/f8fafc/0f172a.png&text=Meeting+Notes+Vision+Benchmark+Run+Pending",
|
||||
"category": "document",
|
||||
"expected_keywords": [
|
||||
"meeting",
|
||||
"notes",
|
||||
"vision"
|
||||
],
|
||||
"ground_truth_ocr": "Meeting Notes Vision Benchmark Run Pending",
|
||||
"expected_structure": {
|
||||
"min_length": 20,
|
||||
"min_sentences": 1,
|
||||
"has_numbers": false
|
||||
}
|
||||
}
|
||||
]
|
||||
]
|
||||
@@ -22,10 +22,12 @@ import argparse
|
||||
import asyncio
|
||||
import base64
|
||||
import json
|
||||
import mimetypes
|
||||
import os
|
||||
import statistics
|
||||
import sys
|
||||
import time
|
||||
import urllib.request
|
||||
from datetime import datetime, timezone
|
||||
from pathlib import Path
|
||||
from typing import Any, Dict, List, Optional
|
||||
@@ -41,12 +43,16 @@ MODELS = {
|
||||
"model_id": "google/gemma-4-27b-it",
|
||||
"display_name": "Gemma 4 27B",
|
||||
"provider": "nous",
|
||||
"fallback_provider": "ollama",
|
||||
"fallback_model_id": "gemma4:latest",
|
||||
"description": "Google's multimodal Gemma 4 model",
|
||||
},
|
||||
"gemini3_flash": {
|
||||
"model_id": "google/gemini-3-flash-preview",
|
||||
"display_name": "Gemini 3 Flash Preview",
|
||||
"provider": "openrouter",
|
||||
"fallback_provider": "gemini",
|
||||
"fallback_model_id": "gemini-2.5-flash",
|
||||
"description": "Current default vision model",
|
||||
},
|
||||
}
|
||||
@@ -84,91 +90,150 @@ async def analyze_with_model(
|
||||
"""
|
||||
import httpx
|
||||
|
||||
def _load_image_bytes_cached() -> tuple[bytes, str]:
|
||||
nonlocal _image_bytes, _mime_type
|
||||
if _image_bytes is not None:
|
||||
return _image_bytes, _mime_type
|
||||
if image_url.startswith(("http://", "https://")):
|
||||
with urllib.request.urlopen(image_url, timeout=30) as resp:
|
||||
_image_bytes = resp.read()
|
||||
_mime_type = resp.headers.get_content_type() or mimetypes.guess_type(image_url)[0] or "image/png"
|
||||
else:
|
||||
path = Path(image_url).expanduser()
|
||||
_image_bytes = path.read_bytes()
|
||||
_mime_type = mimetypes.guess_type(str(path))[0] or "image/png"
|
||||
return _image_bytes, _mime_type
|
||||
|
||||
def _data_url() -> str:
|
||||
image_bytes, mime_type = _load_image_bytes_cached()
|
||||
return f"data:{mime_type};base64,{base64.b64encode(image_bytes).decode()}"
|
||||
|
||||
def _provider_key(provider: str) -> str:
|
||||
if provider == "openrouter":
|
||||
return os.getenv("OPENROUTER_API_KEY", "")
|
||||
if provider == "nous":
|
||||
return os.getenv("NOUS_API_KEY", "") or os.getenv("NOUS_INFERENCE_API_KEY", "")
|
||||
if provider == "gemini":
|
||||
return os.getenv("GEMINI_API_KEY", "") or os.getenv("GOOGLE_API_KEY", "")
|
||||
return os.getenv(f"{provider.upper()}_API_KEY", "")
|
||||
|
||||
provider = model_config["provider"]
|
||||
model_id = model_config["model_id"]
|
||||
candidates = [(provider, model_id)]
|
||||
if model_config.get("fallback_provider") and model_config.get("fallback_model_id"):
|
||||
candidates.append((model_config["fallback_provider"], model_config["fallback_model_id"]))
|
||||
|
||||
# Prepare messages
|
||||
messages = [
|
||||
{
|
||||
"role": "user",
|
||||
"content": [
|
||||
{"type": "text", "text": prompt},
|
||||
{"type": "image_url", "image_url": {"url": image_url}},
|
||||
],
|
||||
}
|
||||
]
|
||||
_image_bytes: Optional[bytes] = None
|
||||
_mime_type = "image/png"
|
||||
failures = []
|
||||
|
||||
# Route to provider
|
||||
if provider == "openrouter":
|
||||
api_url = "https://openrouter.ai/api/v1/chat/completions"
|
||||
api_key = os.getenv("OPENROUTER_API_KEY", "")
|
||||
elif provider == "nous":
|
||||
api_url = "https://inference.nousresearch.com/v1/chat/completions"
|
||||
api_key = os.getenv("NOUS_API_KEY", "") or os.getenv("NOUS_INFERENCE_API_KEY", "")
|
||||
else:
|
||||
api_url = os.getenv(f"{provider.upper()}_API_URL", "")
|
||||
api_key = os.getenv(f"{provider.upper()}_API_KEY", "")
|
||||
for candidate_provider, candidate_model in candidates:
|
||||
api_key = _provider_key(candidate_provider)
|
||||
start = time.perf_counter()
|
||||
try:
|
||||
if candidate_provider in {"openrouter", "nous"}:
|
||||
api_url = (
|
||||
"https://openrouter.ai/api/v1/chat/completions"
|
||||
if candidate_provider == "openrouter"
|
||||
else "https://inference.nousresearch.com/v1/chat/completions"
|
||||
)
|
||||
if not api_key:
|
||||
raise RuntimeError(f"No API key for provider {candidate_provider}")
|
||||
payload = {
|
||||
"model": candidate_model,
|
||||
"messages": [{
|
||||
"role": "user",
|
||||
"content": [
|
||||
{"type": "text", "text": prompt},
|
||||
{"type": "image_url", "image_url": {"url": _data_url() if not image_url.startswith(("http://", "https://")) else image_url}},
|
||||
],
|
||||
}],
|
||||
"max_tokens": 2000,
|
||||
"temperature": 0.1,
|
||||
}
|
||||
headers = {
|
||||
"Authorization": f"Bearer {api_key}",
|
||||
"Content-Type": "application/json",
|
||||
}
|
||||
async with httpx.AsyncClient(timeout=timeout) as client:
|
||||
resp = await client.post(api_url, json=payload, headers=headers)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
analysis = data.get("choices", [{}])[0].get("message", {}).get("content", "")
|
||||
usage = data.get("usage", {})
|
||||
tokens = {
|
||||
"prompt_tokens": usage.get("prompt_tokens", 0),
|
||||
"completion_tokens": usage.get("completion_tokens", 0),
|
||||
"total_tokens": usage.get("total_tokens", 0),
|
||||
}
|
||||
elif candidate_provider == "gemini":
|
||||
if not api_key:
|
||||
raise RuntimeError("No API key for provider gemini")
|
||||
image_bytes, mime_type = _load_image_bytes_cached()
|
||||
api_url = f"https://generativelanguage.googleapis.com/v1beta/models/{candidate_model}:generateContent?key={api_key}"
|
||||
payload = {
|
||||
"contents": [{"parts": [
|
||||
{"text": prompt},
|
||||
{"inline_data": {"mime_type": mime_type, "data": base64.b64encode(image_bytes).decode()}},
|
||||
]}],
|
||||
"generationConfig": {"temperature": 0.1, "maxOutputTokens": 2000},
|
||||
}
|
||||
async with httpx.AsyncClient(timeout=timeout) as client:
|
||||
resp = await client.post(api_url, json=payload)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
parts = data.get("candidates", [{}])[0].get("content", {}).get("parts", [])
|
||||
analysis = "\n".join(part.get("text", "") for part in parts if isinstance(part, dict) and part.get("text"))
|
||||
usage = data.get("usageMetadata", {})
|
||||
tokens = {
|
||||
"prompt_tokens": usage.get("promptTokenCount", 0),
|
||||
"completion_tokens": usage.get("candidatesTokenCount", 0),
|
||||
"total_tokens": usage.get("totalTokenCount", 0),
|
||||
}
|
||||
elif candidate_provider == "ollama":
|
||||
image_bytes, _ = _load_image_bytes_cached()
|
||||
payload = {
|
||||
"model": candidate_model,
|
||||
"stream": False,
|
||||
"messages": [{"role": "user", "content": prompt, "images": [base64.b64encode(image_bytes).decode()]}],
|
||||
"options": {"temperature": 0.1},
|
||||
}
|
||||
async with httpx.AsyncClient(timeout=timeout) as client:
|
||||
resp = await client.post("http://localhost:11434/api/chat", json=payload)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
analysis = data.get("message", {}).get("content", "")
|
||||
tokens = {
|
||||
"prompt_tokens": data.get("prompt_eval_count", 0),
|
||||
"completion_tokens": data.get("eval_count", 0),
|
||||
"total_tokens": (data.get("prompt_eval_count", 0) or 0) + (data.get("eval_count", 0) or 0),
|
||||
}
|
||||
else:
|
||||
raise RuntimeError(f"Unsupported provider {candidate_provider}")
|
||||
|
||||
if not api_key:
|
||||
return {
|
||||
"analysis": "",
|
||||
"latency_ms": 0,
|
||||
"tokens": {},
|
||||
"success": False,
|
||||
"error": f"No API key for provider {provider}",
|
||||
}
|
||||
latency_ms = (time.perf_counter() - start) * 1000
|
||||
return {
|
||||
"analysis": analysis,
|
||||
"latency_ms": round(latency_ms, 1),
|
||||
"tokens": tokens,
|
||||
"success": True,
|
||||
"error": "",
|
||||
"provider_used": candidate_provider,
|
||||
"model_used": candidate_model,
|
||||
}
|
||||
except Exception as e:
|
||||
failures.append(f"{candidate_provider}:{candidate_model} => {e}")
|
||||
|
||||
headers = {
|
||||
"Authorization": f"Bearer {api_key}",
|
||||
"Content-Type": "application/json",
|
||||
return {
|
||||
"analysis": "",
|
||||
"latency_ms": 0,
|
||||
"tokens": {},
|
||||
"success": False,
|
||||
"error": " | ".join(failures) if failures else "No runs",
|
||||
"provider_used": candidates[-1][0] if candidates else provider,
|
||||
"model_used": candidates[-1][1] if candidates else model_id,
|
||||
}
|
||||
|
||||
payload = {
|
||||
"model": model_id,
|
||||
"messages": messages,
|
||||
"max_tokens": 2000,
|
||||
"temperature": 0.1,
|
||||
}
|
||||
|
||||
start = time.perf_counter()
|
||||
try:
|
||||
async with httpx.AsyncClient(timeout=timeout) as client:
|
||||
resp = await client.post(api_url, json=payload, headers=headers)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
|
||||
latency_ms = (time.perf_counter() - start) * 1000
|
||||
|
||||
analysis = ""
|
||||
choices = data.get("choices", [])
|
||||
if choices:
|
||||
msg = choices[0].get("message", {})
|
||||
analysis = msg.get("content", "")
|
||||
|
||||
usage = data.get("usage", {})
|
||||
tokens = {
|
||||
"prompt_tokens": usage.get("prompt_tokens", 0),
|
||||
"completion_tokens": usage.get("completion_tokens", 0),
|
||||
"total_tokens": usage.get("total_tokens", 0),
|
||||
}
|
||||
|
||||
return {
|
||||
"analysis": analysis,
|
||||
"latency_ms": round(latency_ms, 1),
|
||||
"tokens": tokens,
|
||||
"success": True,
|
||||
"error": "",
|
||||
}
|
||||
|
||||
except Exception as e:
|
||||
return {
|
||||
"analysis": "",
|
||||
"latency_ms": round((time.perf_counter() - start) * 1000, 1),
|
||||
"tokens": {},
|
||||
"success": False,
|
||||
"error": str(e),
|
||||
}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Evaluation metrics
|
||||
@@ -398,7 +463,13 @@ def aggregate_results(results: List[dict], models: dict) -> dict:
|
||||
failed = [r[model_name] for r in results if not r[model_name]["success"]]
|
||||
|
||||
if not model_results:
|
||||
summary[model_name] = {"success_rate": 0, "error": "All runs failed"}
|
||||
summary[model_name] = {
|
||||
"success_rate": 0,
|
||||
"error": "All runs failed",
|
||||
"total_runs": 0,
|
||||
"total_failures": len(failed),
|
||||
"failure_examples": sorted({f.get("error", "unknown failure") for f in failed})[:3],
|
||||
}
|
||||
continue
|
||||
|
||||
latencies = [r["avg_latency_ms"] for r in model_results]
|
||||
@@ -410,6 +481,7 @@ def aggregate_results(results: List[dict], models: dict) -> dict:
|
||||
"success_rate": round(len(model_results) / (len(model_results) + len(failed)), 4),
|
||||
"total_runs": len(model_results),
|
||||
"total_failures": len(failed),
|
||||
"failure_examples": sorted({f.get("error", "unknown failure") for f in failed})[:3],
|
||||
"latency": {
|
||||
"mean_ms": round(statistics.mean(latencies), 1),
|
||||
"median_ms": round(statistics.median(latencies), 1),
|
||||
@@ -495,6 +567,23 @@ def to_markdown(report: dict) -> str:
|
||||
f"| {mname} | {tok['mean_total']:.0f} | {tok['total_used']} |"
|
||||
)
|
||||
|
||||
lines += ["", "## Failure Modes", ""]
|
||||
had_failures = False
|
||||
for mkey, mname in config["models"].items():
|
||||
model_summary = summary.get(mkey, {})
|
||||
failure_examples = model_summary.get("failure_examples", [])
|
||||
if not failure_examples and not model_summary.get("error"):
|
||||
continue
|
||||
had_failures = True
|
||||
lines.append(f"### {mname}")
|
||||
if model_summary.get("error"):
|
||||
lines.append(f"- Summary: {model_summary['error']}")
|
||||
for err in failure_examples:
|
||||
lines.append(f"- {err}")
|
||||
lines.append("")
|
||||
if not had_failures:
|
||||
lines.append("- No provider/runtime failures recorded.")
|
||||
|
||||
# Verdict
|
||||
lines += ["", "## Verdict", ""]
|
||||
|
||||
@@ -516,8 +605,12 @@ def to_markdown(report: dict) -> str:
|
||||
|
||||
if best_model:
|
||||
lines.append(f"**Best overall: {best_model}** (composite score: {best_score:.1%})")
|
||||
lines.append("")
|
||||
lines.append("Recommendation: keep the best-performing Gemma/Gemini lane from this run and only switch if repeated runs disagree.")
|
||||
else:
|
||||
lines.append("No clear winner — insufficient data.")
|
||||
lines.append("Benchmark blocked or insufficient data for a trustworthy winner.")
|
||||
lines.append("")
|
||||
lines.append("Recommendation: repair provider/runtime availability, rerun the benchmark, and keep the current implementation unchanged until comparative results exist.")
|
||||
|
||||
return "\n".join(lines)
|
||||
|
||||
@@ -528,44 +621,124 @@ def to_markdown(report: dict) -> str:
|
||||
|
||||
|
||||
def generate_sample_dataset() -> List[dict]:
|
||||
"""Generate a sample test dataset with diverse public images.
|
||||
"""Generate a larger benchmark dataset aligned with issue #817.
|
||||
|
||||
Returns list of test image definitions.
|
||||
Returns 50+ images across screenshots, diagrams, photos, OCR, charts,
|
||||
and document-like images so the harness matches the issue contract.
|
||||
"""
|
||||
return [
|
||||
# Screenshots
|
||||
{
|
||||
"id": "screenshot_github",
|
||||
"url": "https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png",
|
||||
dataset: List[dict] = []
|
||||
|
||||
screenshots = [
|
||||
("github_mark", "https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png", ["github", "logo", "mark"]),
|
||||
("github_social", "https://github.githubassets.com/images/modules/site/social-cards.png", ["github", "page", "web"]),
|
||||
("github_code_search", "https://github.githubassets.com/images/modules/site/features-code-search.png", ["search", "code", "feature"]),
|
||||
("terminal_capture", "https://raw.githubusercontent.com/nicehash/nicehash-quick-start/main/images/nicehash-terminal.png", ["terminal", "command", "output"]),
|
||||
("http_404", "https://http.cat/404.jpg", ["404", "error", "cat"]),
|
||||
("dummy_cli_01", "https://dummyimage.com/1280x720/111827/f9fafb.png&text=Hermes+CLI+Session+01", ["hermes", "cli", "session"]),
|
||||
("dummy_cli_02", "https://dummyimage.com/1280x720/0f172a/e2e8f0.png&text=Prompt+Cache+Dashboard", ["prompt", "cache", "dashboard"]),
|
||||
("dummy_ui_01", "https://dummyimage.com/1280x720/1f2937/f3f4f6.png&text=Settings+Panel+Voice+Mode", ["settings", "voice", "mode"]),
|
||||
("dummy_ui_02", "https://dummyimage.com/1280x720/334155/f8fafc.png&text=Browser+Vision+Preview", ["browser", "vision", "preview"]),
|
||||
("dummy_ui_03", "https://dummyimage.com/1280x720/111827/ffffff.png&text=Tool+Call+Inspector", ["tool", "call", "inspector"]),
|
||||
]
|
||||
for ident, url, keywords in screenshots:
|
||||
dataset.append({
|
||||
"id": f"screenshot_{ident}",
|
||||
"url": url,
|
||||
"category": "screenshot",
|
||||
"expected_keywords": ["github", "logo", "octocat"],
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2},
|
||||
},
|
||||
# Diagrams
|
||||
{
|
||||
"id": "diagram_architecture",
|
||||
"url": "https://mermaid.ink/img/pako:eNp9kMtOwzAQRX_F8hKpJbhJFVJBi1QJiMWCG8eZNsGJLdlOiqIid5RdufiHnZRA7GbuzJwZe4ZGH2SCBPYUwgxoQKvJnCR2YY0F5YBdJJkD4uX0oXB6PnF3U4zCWcWdW3FqOwGvCKkBmHKSTB2gJeRrLTeJLfJdJKkBGYf9P1sTNdUXVJqY3YNJK7xLVwR0mxJFU6rCgEKnhSGIL2Eq8BdEERAX0OGwEiVQ1R0MaNFR8QfqKxmHigbX8VLjDz_Q0L8Wc_qPxDw",
|
||||
"expected_keywords": keywords,
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": False},
|
||||
})
|
||||
|
||||
diagrams = [
|
||||
("flow_a", "https://dummyimage.com/1200x800/f8fafc/0f172a.png&text=Flowchart+API+Gateway+Queue+Worker", ["flowchart", "api", "worker"]),
|
||||
("flow_b", "https://dummyimage.com/1200x800/f1f5f9/0f172a.png&text=Architecture+Diagram+Database+Cache+Client", ["architecture", "diagram", "cache"]),
|
||||
("uml_a", "https://dummyimage.com/1200x800/e2e8f0/0f172a.png&text=Class+Diagram+User+Session+Message", ["class", "diagram", "session"]),
|
||||
("uml_b", "https://dummyimage.com/1200x800/cbd5e1/0f172a.png&text=Sequence+Diagram+Request+Response", ["sequence", "diagram", "response"]),
|
||||
("network_a", "https://dummyimage.com/1200x800/ffffff/111827.png&text=Network+Nodes+Edges+Router", ["network", "node", "router"]),
|
||||
("network_b", "https://dummyimage.com/1200x800/ffffff/1e293b.png&text=Service+Mesh+Proxy+Auth", ["service", "mesh", "auth"]),
|
||||
("state_machine", "https://dummyimage.com/1200x800/f8fafc/334155.png&text=State+Machine+Idle+Run+Stop", ["state", "machine", "idle"]),
|
||||
("mind_map", "https://dummyimage.com/1200x800/fefce8/1f2937.png&text=Mind+Map+Memory+Recall+Tools", ["mind", "memory", "tools"]),
|
||||
("pipeline", "https://dummyimage.com/1200x800/ecfeff/155e75.png&text=Pipeline+Ingest+Rank+Summarize", ["pipeline", "ingest", "summarize"]),
|
||||
("org_chart", "https://dummyimage.com/1200x800/fdf2f8/831843.png&text=Org+Chart+Lead+Review+Ops", ["org", "chart", "review"]),
|
||||
]
|
||||
for ident, url, keywords in diagrams:
|
||||
dataset.append({
|
||||
"id": f"diagram_{ident}",
|
||||
"url": url,
|
||||
"category": "diagram",
|
||||
"expected_keywords": ["architecture", "component", "service"],
|
||||
"expected_structure": {"min_length": 100, "min_sentences": 3},
|
||||
},
|
||||
# Photos
|
||||
{
|
||||
"id": "photo_nature",
|
||||
"url": "https://picsum.photos/seed/bench1/400/300",
|
||||
"expected_keywords": keywords,
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": False},
|
||||
})
|
||||
|
||||
for idx in range(1, 11):
|
||||
dataset.append({
|
||||
"id": f"photo_random_{idx:02d}",
|
||||
"url": f"https://picsum.photos/seed/vision-bench-{idx}/640/480",
|
||||
"category": "photo",
|
||||
"expected_keywords": [],
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1},
|
||||
},
|
||||
# Charts
|
||||
{
|
||||
"id": "chart_bar",
|
||||
"url": "https://quickchart.io/chart?c={type:'bar',data:{labels:['Q1','Q2','Q3','Q4'],datasets:[{label:'Users',data:[50,60,70,80]}]}}",
|
||||
"category": "chart",
|
||||
"expected_keywords": ["bar", "chart", "data"],
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2},
|
||||
},
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 30, "min_sentences": 1, "has_numbers": False},
|
||||
})
|
||||
|
||||
charts = [
|
||||
("bar_quarterly", "https://quickchart.io/chart?c={type:'bar',data:{labels:['Q1','Q2','Q3','Q4'],datasets:[{label:'Revenue',data:[100,150,200,250]}]}}", ["bar", "chart", "revenue"]),
|
||||
("pie_market", "https://quickchart.io/chart?c={type:'pie',data:{labels:['A','B','C'],datasets:[{data:[30,50,20]}]}}", ["pie", "chart", "percentage"]),
|
||||
("line_temp", "https://quickchart.io/chart?c={type:'line',data:{labels:['Jan','Feb','Mar','Apr'],datasets:[{label:'Temperature',data:[5,8,12,18]}]}}", ["line", "chart", "temperature"]),
|
||||
("radar_skill", "https://quickchart.io/chart?c={type:'radar',data:{labels:['Speed','Power','Defense','Magic'],datasets:[{label:'Hero',data:[80,60,70,90]}]}}", ["radar", "chart", "skill"]),
|
||||
("stacked_cloud", "https://quickchart.io/chart?c={type:'bar',data:{labels:['2022','2023','2024'],datasets:[{label:'Cloud',data:[100,150,200]},{label:'On-prem',data:[200,180,160]}]},options:{scales:{x:{stacked:true},y:{stacked:true}}}}", ["stacked", "bar", "chart"]),
|
||||
("area_growth", "https://quickchart.io/chart?c={type:'line',data:{labels:['W1','W2','W3','W4'],datasets:[{label:'Growth',data:[10,15,18,24],fill:true}]}}", ["line", "growth", "chart"]),
|
||||
("scatter_eval", "https://quickchart.io/chart?c={type:'scatter',data:{datasets:[{label:'Runs',data:[{x:1,y:70},{x:2,y:75},{x:3,y:82}]}]}}", ["scatter", "chart", "runs"]),
|
||||
("horizontal_bar", "https://quickchart.io/chart?c={type:'bar',data:{labels:['UI','OCR','Docs'],datasets:[{label:'Score',data:[88,76,91]}]},options:{indexAxis:'y'}}", ["bar", "score", "ocr"]),
|
||||
("bubble_usage", "https://quickchart.io/chart?c={type:'bubble',data:{datasets:[{label:'Latency',data:[{x:1,y:120,r:8},{x:2,y:95,r:6},{x:3,y:180,r:10}]}]}}", ["bubble", "latency", "chart"]),
|
||||
("doughnut_devices", "https://quickchart.io/chart?c={type:'doughnut',data:{labels:['Desktop','Mobile','Tablet'],datasets:[{data:[60,30,10]}]}}", ["doughnut", "chart", "device"]),
|
||||
]
|
||||
for ident, url, keywords in charts:
|
||||
dataset.append({
|
||||
"id": f"chart_{ident}",
|
||||
"url": url,
|
||||
"category": "chart",
|
||||
"expected_keywords": keywords,
|
||||
"ground_truth_ocr": "",
|
||||
"expected_structure": {"min_length": 50, "min_sentences": 2, "has_numbers": True},
|
||||
})
|
||||
|
||||
ocr_texts = [
|
||||
"Hermes OCR Alpha 01",
|
||||
"Prompt Cache Hit 87%",
|
||||
"Session 42 Ready",
|
||||
"Latency 118 ms",
|
||||
"Voice Mode Enabled",
|
||||
]
|
||||
for idx, text in enumerate(ocr_texts, start=1):
|
||||
dataset.append({
|
||||
"id": f"ocr_text_{idx:02d}",
|
||||
"url": f"https://dummyimage.com/1200x320/ffffff/000000.png&text={text.replace(' ', '+')}",
|
||||
"category": "ocr",
|
||||
"expected_keywords": text.lower().split()[:2],
|
||||
"ground_truth_ocr": text,
|
||||
"expected_structure": {"min_length": 10, "min_sentences": 1, "has_numbers": any(ch.isdigit() for ch in text)},
|
||||
})
|
||||
|
||||
documents = [
|
||||
"Invoice 1001 Total 42 Due 2026-04-22",
|
||||
"Form A Name Alice Status Approved",
|
||||
"Report Memory Recall Score 91 Percent",
|
||||
"Checklist Crisis Escalation Call 988 Now",
|
||||
"Meeting Notes Vision Benchmark Run Pending",
|
||||
]
|
||||
for idx, text in enumerate(documents, start=1):
|
||||
dataset.append({
|
||||
"id": f"document_text_{idx:02d}",
|
||||
"url": f"https://dummyimage.com/1400x900/f8fafc/0f172a.png&text={text.replace(' ', '+')}",
|
||||
"category": "document",
|
||||
"expected_keywords": text.lower().split()[:3],
|
||||
"ground_truth_ocr": text,
|
||||
"expected_structure": {"min_length": 20, "min_sentences": 1, "has_numbers": any(ch.isdigit() for ch in text)},
|
||||
})
|
||||
|
||||
return dataset
|
||||
|
||||
|
||||
def load_dataset(path: str) -> List[dict]:
|
||||
@@ -585,7 +758,9 @@ async def main():
|
||||
parser.add_argument("--url", help="Single image URL to test")
|
||||
parser.add_argument("--category", default="photo", help="Category for single URL")
|
||||
parser.add_argument("--output", default=None, help="Output JSON file")
|
||||
parser.add_argument("--markdown-output", default=None, help="Optional markdown report output path")
|
||||
parser.add_argument("--runs", type=int, default=1, help="Runs per model per image")
|
||||
parser.add_argument("--limit", type=int, default=0, help="Limit to the first N images for smoke runs")
|
||||
parser.add_argument("--models", nargs="+", default=None,
|
||||
help="Models to test (default: all)")
|
||||
parser.add_argument("--markdown", action="store_true", help="Output markdown report")
|
||||
@@ -617,9 +792,14 @@ async def main():
|
||||
print("ERROR: Provide --images or --url")
|
||||
sys.exit(1)
|
||||
|
||||
if args.limit and args.limit > 0:
|
||||
images = images[:args.limit]
|
||||
|
||||
# Run benchmark
|
||||
report = await run_benchmark_suite(images, selected, args.runs)
|
||||
|
||||
markdown_report = to_markdown(report)
|
||||
|
||||
# Output
|
||||
if args.output:
|
||||
os.makedirs(os.path.dirname(args.output) or ".", exist_ok=True)
|
||||
@@ -627,8 +807,14 @@ async def main():
|
||||
json.dump(report, f, indent=2)
|
||||
print(f"\nResults saved to {args.output}")
|
||||
|
||||
if args.markdown_output:
|
||||
os.makedirs(os.path.dirname(args.markdown_output) or ".", exist_ok=True)
|
||||
with open(args.markdown_output, "w", encoding="utf-8") as f:
|
||||
f.write(markdown_report)
|
||||
print(f"Markdown report saved to {args.markdown_output}")
|
||||
|
||||
if args.markdown or not args.output:
|
||||
print("\n" + to_markdown(report))
|
||||
print("\n" + markdown_report)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
67
metrics/vision-benchmark-smoke-2026-04-22.json
Normal file
67
metrics/vision-benchmark-smoke-2026-04-22.json
Normal file
@@ -0,0 +1,67 @@
|
||||
{
|
||||
"generated_at": "2026-04-22T16:21:56.271426+00:00",
|
||||
"config": {
|
||||
"total_images": 2,
|
||||
"runs_per_model": 1,
|
||||
"models": {
|
||||
"gemma4": "Gemma 4 27B",
|
||||
"gemini3_flash": "Gemini 3 Flash Preview"
|
||||
}
|
||||
},
|
||||
"results": [
|
||||
{
|
||||
"gemma4": {
|
||||
"success": false,
|
||||
"error": "nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => Server error '500 Internal Server Error' for url 'http://localhost:11434/api/chat'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/500",
|
||||
"runs": 0,
|
||||
"errors": 1
|
||||
},
|
||||
"gemini3_flash": {
|
||||
"success": false,
|
||||
"error": "openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => Client error '429 Too Many Requests' for url 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=AIzaSyAmIctJQG_b4VKV1sMLebBnouq6yCckEf0'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/429",
|
||||
"runs": 0,
|
||||
"errors": 1
|
||||
},
|
||||
"image_id": "screenshot_github_mark",
|
||||
"category": "screenshot"
|
||||
},
|
||||
{
|
||||
"gemma4": {
|
||||
"success": false,
|
||||
"error": "nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => HTTP Error 404: Not Found",
|
||||
"runs": 0,
|
||||
"errors": 1
|
||||
},
|
||||
"gemini3_flash": {
|
||||
"success": false,
|
||||
"error": "openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => HTTP Error 404: Not Found",
|
||||
"runs": 0,
|
||||
"errors": 1
|
||||
},
|
||||
"image_id": "screenshot_github_social",
|
||||
"category": "screenshot"
|
||||
}
|
||||
],
|
||||
"summary": {
|
||||
"gemma4": {
|
||||
"success_rate": 0,
|
||||
"error": "All runs failed",
|
||||
"total_runs": 0,
|
||||
"total_failures": 2,
|
||||
"failure_examples": [
|
||||
"nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => HTTP Error 404: Not Found",
|
||||
"nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => Server error '500 Internal Server Error' for url 'http://localhost:11434/api/chat'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/500"
|
||||
]
|
||||
},
|
||||
"gemini3_flash": {
|
||||
"success_rate": 0,
|
||||
"error": "All runs failed",
|
||||
"total_runs": 0,
|
||||
"total_failures": 2,
|
||||
"failure_examples": [
|
||||
"openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => Client error '429 Too Many Requests' for url 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=AIzaSyAmIctJQG_b4VKV1sMLebBnouq6yCckEf0'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/429",
|
||||
"openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => HTTP Error 404: Not Found"
|
||||
]
|
||||
}
|
||||
}
|
||||
}
|
||||
44
metrics/vision-benchmark-smoke-2026-04-22.md
Normal file
44
metrics/vision-benchmark-smoke-2026-04-22.md
Normal file
@@ -0,0 +1,44 @@
|
||||
# Vision Benchmark Report
|
||||
|
||||
Generated: 2026-04-22T16:21
|
||||
Images tested: 2
|
||||
Runs per model: 1
|
||||
Models: Gemma 4 27B, Gemini 3 Flash Preview
|
||||
|
||||
## Latency Comparison
|
||||
|
||||
| Model | Mean (ms) | Median | P95 | Std Dev |
|
||||
|-------|-----------|--------|-----|---------|
|
||||
|
||||
## Accuracy Comparison
|
||||
|
||||
| Model | OCR Accuracy | Keyword Coverage | Success Rate |
|
||||
|-------|-------------|-----------------|--------------|
|
||||
|
||||
## Token Usage
|
||||
|
||||
| Model | Mean Tokens/Image | Total Tokens |
|
||||
|-------|------------------|--------------|
|
||||
|
||||
## Failure Modes
|
||||
|
||||
### Gemma 4 27B
|
||||
- Summary: All runs failed
|
||||
- nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => HTTP Error 404: Not Found
|
||||
- nous:google/gemma-4-27b-it => No API key for provider nous | ollama:gemma4:latest => Server error '500 Internal Server Error' for url 'http://localhost:11434/api/chat'
|
||||
For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/500
|
||||
|
||||
### Gemini 3 Flash Preview
|
||||
- Summary: All runs failed
|
||||
- openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'
|
||||
For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => Client error '429 Too Many Requests' for url 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=AIzaSyAmIctJQG_b4VKV1sMLebBnouq6yCckEf0'
|
||||
For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/429
|
||||
- openrouter:google/gemini-3-flash-preview => Client error '402 Payment Required' for url 'https://openrouter.ai/api/v1/chat/completions'
|
||||
For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402 | gemini:gemini-2.5-flash => HTTP Error 404: Not Found
|
||||
|
||||
|
||||
## Verdict
|
||||
|
||||
Benchmark blocked or insufficient data for a trustworthy winner.
|
||||
|
||||
Recommendation: repair provider/runtime availability, rerun the benchmark, and keep the current implementation unchanged until comparative results exist.
|
||||
@@ -356,57 +356,44 @@ class HolographicMemoryProvider(MemoryProvider):
|
||||
# -- Auto-extraction (on_session_end) ------------------------------------
|
||||
|
||||
def _auto_extract_facts(self, messages: list) -> None:
|
||||
from agent.session_compactor import evaluate_extraction_quality, extract_facts_from_messages
|
||||
|
||||
def _store_category(category: str) -> str:
|
||||
if category.startswith("user_pref"):
|
||||
return "user_pref"
|
||||
if category.startswith("project"):
|
||||
return "project"
|
||||
if category.startswith("tool"):
|
||||
return "tool"
|
||||
return "general"
|
||||
|
||||
facts = extract_facts_from_messages(messages)
|
||||
if not facts:
|
||||
return
|
||||
_PREF_PATTERNS = [
|
||||
re.compile(r'\bI\s+(?:prefer|like|love|use|want|need)\s+(.+)', re.IGNORECASE),
|
||||
re.compile(r'\bmy\s+(?:favorite|preferred|default)\s+\w+\s+is\s+(.+)', re.IGNORECASE),
|
||||
re.compile(r'\bI\s+(?:always|never|usually)\s+(.+)', re.IGNORECASE),
|
||||
]
|
||||
_DECISION_PATTERNS = [
|
||||
re.compile(r'\bwe\s+(?:decided|agreed|chose)\s+(?:to\s+)?(.+)', re.IGNORECASE),
|
||||
re.compile(r'\bthe\s+project\s+(?:uses|needs|requires)\s+(.+)', re.IGNORECASE),
|
||||
]
|
||||
|
||||
extracted = 0
|
||||
for fact in facts:
|
||||
try:
|
||||
metadata = dict(fact.metadata)
|
||||
metadata.setdefault("relation", fact.relation)
|
||||
metadata.setdefault("value", fact.content)
|
||||
metadata.setdefault("provenance", [fact.provenance])
|
||||
metadata.setdefault("evidence", list(fact.evidence))
|
||||
metadata.setdefault("observation_count", len(fact.evidence))
|
||||
metadata.setdefault("duplicate_count", max(0, len(fact.evidence) - 1))
|
||||
self._store.add_fact(
|
||||
fact.content,
|
||||
category=_store_category(fact.category),
|
||||
tags=",".join(filter(None, [fact.entity, fact.relation, fact.status])),
|
||||
canonical_key=fact.canonical_key,
|
||||
metadata=metadata,
|
||||
confidence=fact.confidence,
|
||||
source_role=fact.source_role,
|
||||
source_turn=fact.source_turn,
|
||||
observed_at=fact.observed_at,
|
||||
contradiction_group=fact.contradiction_group,
|
||||
status=fact.status,
|
||||
)
|
||||
extracted += 1
|
||||
except Exception as exc:
|
||||
logger.debug("Structured auto-extract failed for %s: %s", fact.canonical_key, exc)
|
||||
for msg in messages:
|
||||
if msg.get("role") != "user":
|
||||
continue
|
||||
content = msg.get("content", "")
|
||||
if not isinstance(content, str) or len(content) < 10:
|
||||
continue
|
||||
|
||||
for pattern in _PREF_PATTERNS:
|
||||
if pattern.search(content):
|
||||
try:
|
||||
self._store.add_fact(content[:400], category="user_pref")
|
||||
extracted += 1
|
||||
except Exception:
|
||||
pass
|
||||
break
|
||||
|
||||
for pattern in _DECISION_PATTERNS:
|
||||
if pattern.search(content):
|
||||
try:
|
||||
self._store.add_fact(content[:400], category="project")
|
||||
extracted += 1
|
||||
except Exception:
|
||||
pass
|
||||
break
|
||||
|
||||
if extracted:
|
||||
metrics = evaluate_extraction_quality(messages)
|
||||
logger.info(
|
||||
"Auto-extracted %d structured facts from conversation (raw=%d normalized=%d contradictions=%d)",
|
||||
extracted,
|
||||
metrics["raw_candidates"],
|
||||
metrics["normalized_facts"],
|
||||
metrics["contradiction_groups"],
|
||||
)
|
||||
logger.info("Auto-extracted %d facts from conversation", extracted)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
@@ -3,7 +3,6 @@ SQLite-backed fact store with entity resolution and trust scoring.
|
||||
Single-user Hermes memory store plugin.
|
||||
"""
|
||||
|
||||
import json
|
||||
import re
|
||||
import sqlite3
|
||||
import threading
|
||||
@@ -16,24 +15,16 @@ except ImportError:
|
||||
|
||||
_SCHEMA = """
|
||||
CREATE TABLE IF NOT EXISTS facts (
|
||||
fact_id INTEGER PRIMARY KEY AUTOINCREMENT,
|
||||
content TEXT NOT NULL UNIQUE,
|
||||
category TEXT DEFAULT 'general',
|
||||
tags TEXT DEFAULT '',
|
||||
trust_score REAL DEFAULT 0.5,
|
||||
retrieval_count INTEGER DEFAULT 0,
|
||||
helpful_count INTEGER DEFAULT 0,
|
||||
canonical_key TEXT DEFAULT '',
|
||||
metadata_json TEXT DEFAULT '{}',
|
||||
confidence REAL DEFAULT 0.5,
|
||||
source_role TEXT DEFAULT '',
|
||||
source_turn INTEGER DEFAULT -1,
|
||||
observed_at TEXT DEFAULT '',
|
||||
contradiction_group TEXT DEFAULT '',
|
||||
status TEXT DEFAULT 'active',
|
||||
created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
|
||||
updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
|
||||
hrr_vector BLOB
|
||||
fact_id INTEGER PRIMARY KEY AUTOINCREMENT,
|
||||
content TEXT NOT NULL UNIQUE,
|
||||
category TEXT DEFAULT 'general',
|
||||
tags TEXT DEFAULT '',
|
||||
trust_score REAL DEFAULT 0.5,
|
||||
retrieval_count INTEGER DEFAULT 0,
|
||||
helpful_count INTEGER DEFAULT 0,
|
||||
created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
|
||||
updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
|
||||
hrr_vector BLOB
|
||||
);
|
||||
|
||||
CREATE TABLE IF NOT EXISTS entities (
|
||||
@@ -50,11 +41,9 @@ CREATE TABLE IF NOT EXISTS fact_entities (
|
||||
PRIMARY KEY (fact_id, entity_id)
|
||||
);
|
||||
|
||||
CREATE INDEX IF NOT EXISTS idx_facts_trust ON facts(trust_score DESC);
|
||||
CREATE INDEX IF NOT EXISTS idx_facts_category ON facts(category);
|
||||
CREATE INDEX IF NOT EXISTS idx_facts_canonical_key ON facts(canonical_key);
|
||||
CREATE INDEX IF NOT EXISTS idx_facts_contradiction_group ON facts(contradiction_group);
|
||||
CREATE INDEX IF NOT EXISTS idx_entities_name ON entities(name);
|
||||
CREATE INDEX IF NOT EXISTS idx_facts_trust ON facts(trust_score DESC);
|
||||
CREATE INDEX IF NOT EXISTS idx_facts_category ON facts(category);
|
||||
CREATE INDEX IF NOT EXISTS idx_entities_name ON entities(name);
|
||||
|
||||
CREATE VIRTUAL TABLE IF NOT EXISTS facts_fts
|
||||
USING fts5(content, tags, content=facts, content_rowid=fact_id);
|
||||
@@ -140,23 +129,10 @@ class MemoryStore:
|
||||
"""Create tables, indexes, and triggers if they do not exist. Enable WAL mode."""
|
||||
self._conn.execute("PRAGMA journal_mode=WAL")
|
||||
self._conn.executescript(_SCHEMA)
|
||||
# Migrate: add hrr_vector column if missing (safe for existing databases)
|
||||
columns = {row[1] for row in self._conn.execute("PRAGMA table_info(facts)").fetchall()}
|
||||
migrations = {
|
||||
"hrr_vector": "ALTER TABLE facts ADD COLUMN hrr_vector BLOB",
|
||||
"canonical_key": "ALTER TABLE facts ADD COLUMN canonical_key TEXT DEFAULT ''",
|
||||
"metadata_json": "ALTER TABLE facts ADD COLUMN metadata_json TEXT DEFAULT '{}'",
|
||||
"confidence": "ALTER TABLE facts ADD COLUMN confidence REAL DEFAULT 0.5",
|
||||
"source_role": "ALTER TABLE facts ADD COLUMN source_role TEXT DEFAULT ''",
|
||||
"source_turn": "ALTER TABLE facts ADD COLUMN source_turn INTEGER DEFAULT -1",
|
||||
"observed_at": "ALTER TABLE facts ADD COLUMN observed_at TEXT DEFAULT ''",
|
||||
"contradiction_group": "ALTER TABLE facts ADD COLUMN contradiction_group TEXT DEFAULT ''",
|
||||
"status": "ALTER TABLE facts ADD COLUMN status TEXT DEFAULT 'active'",
|
||||
}
|
||||
for column, ddl in migrations.items():
|
||||
if column not in columns:
|
||||
self._conn.execute(ddl)
|
||||
self._conn.execute("CREATE INDEX IF NOT EXISTS idx_facts_canonical_key ON facts(canonical_key)")
|
||||
self._conn.execute("CREATE INDEX IF NOT EXISTS idx_facts_contradiction_group ON facts(contradiction_group)")
|
||||
if "hrr_vector" not in columns:
|
||||
self._conn.execute("ALTER TABLE facts ADD COLUMN hrr_vector BLOB")
|
||||
self._conn.commit()
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
@@ -168,148 +144,41 @@ class MemoryStore:
|
||||
content: str,
|
||||
category: str = "general",
|
||||
tags: str = "",
|
||||
*,
|
||||
canonical_key: str = "",
|
||||
metadata: dict | None = None,
|
||||
confidence: float | None = None,
|
||||
source_role: str = "",
|
||||
source_turn: int = -1,
|
||||
observed_at: str = "",
|
||||
contradiction_group: str = "",
|
||||
status: str = "active",
|
||||
) -> int:
|
||||
"""Insert a fact and return its fact_id.
|
||||
|
||||
Exact duplicates are deduplicated by content. Near-duplicates are
|
||||
normalized by canonical_key, with provenance/evidence merged into the
|
||||
existing row. Contradictions sharing the same contradiction_group remain
|
||||
stored as separate rows and are marked inspectably.
|
||||
Deduplicates by content (UNIQUE constraint). On duplicate, returns
|
||||
the existing fact_id without modifying the row. Extracts entities from
|
||||
the content and links them to the fact.
|
||||
"""
|
||||
with self._lock:
|
||||
content = content.strip()
|
||||
if not content:
|
||||
raise ValueError("content must not be empty")
|
||||
|
||||
metadata = dict(metadata or {})
|
||||
canonical_key = canonical_key.strip()
|
||||
contradiction_group = contradiction_group.strip()
|
||||
observed_at = observed_at.strip()
|
||||
status = status or "active"
|
||||
trust_score = self.default_trust if confidence is None else _clamp_trust(confidence)
|
||||
metadata_json = json.dumps(metadata, sort_keys=True)
|
||||
|
||||
if canonical_key:
|
||||
existing = self._conn.execute(
|
||||
"SELECT fact_id, metadata_json, trust_score, confidence, observed_at FROM facts WHERE canonical_key = ?",
|
||||
(canonical_key,),
|
||||
).fetchone()
|
||||
if existing is not None:
|
||||
merged_metadata = self._merge_metadata(existing["metadata_json"], metadata)
|
||||
merged_trust = max(float(existing["trust_score"]), trust_score)
|
||||
merged_observed_at = existing["observed_at"] or observed_at
|
||||
if observed_at and merged_observed_at:
|
||||
merged_observed_at = min(merged_observed_at, observed_at)
|
||||
elif observed_at:
|
||||
merged_observed_at = observed_at
|
||||
self._conn.execute(
|
||||
"""
|
||||
UPDATE facts
|
||||
SET metadata_json = ?,
|
||||
trust_score = ?,
|
||||
confidence = ?,
|
||||
observed_at = ?,
|
||||
updated_at = CURRENT_TIMESTAMP
|
||||
WHERE fact_id = ?
|
||||
""",
|
||||
(
|
||||
json.dumps(merged_metadata, sort_keys=True),
|
||||
merged_trust,
|
||||
max(float(existing["confidence"] or 0.0), confidence or trust_score),
|
||||
merged_observed_at,
|
||||
existing["fact_id"],
|
||||
),
|
||||
)
|
||||
self._conn.commit()
|
||||
return int(existing["fact_id"])
|
||||
|
||||
contradiction_rows = []
|
||||
if contradiction_group:
|
||||
contradiction_rows = self._conn.execute(
|
||||
"""
|
||||
SELECT fact_id, canonical_key, metadata_json
|
||||
FROM facts
|
||||
WHERE contradiction_group = ?
|
||||
AND canonical_key != ?
|
||||
""",
|
||||
(contradiction_group, canonical_key),
|
||||
).fetchall()
|
||||
if contradiction_rows:
|
||||
status = "contradiction"
|
||||
metadata = dict(metadata)
|
||||
metadata["status"] = "contradiction"
|
||||
metadata["contradiction_group"] = contradiction_group
|
||||
metadata["contradiction_keys"] = sorted(
|
||||
{
|
||||
canonical_key,
|
||||
*[str(row["canonical_key"]) for row in contradiction_rows if row["canonical_key"]],
|
||||
}
|
||||
- {""}
|
||||
)
|
||||
metadata_json = json.dumps(metadata, sort_keys=True)
|
||||
|
||||
try:
|
||||
cur = self._conn.execute(
|
||||
"""
|
||||
INSERT INTO facts (
|
||||
content,
|
||||
category,
|
||||
tags,
|
||||
trust_score,
|
||||
canonical_key,
|
||||
metadata_json,
|
||||
confidence,
|
||||
source_role,
|
||||
source_turn,
|
||||
observed_at,
|
||||
contradiction_group,
|
||||
status
|
||||
)
|
||||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
|
||||
INSERT INTO facts (content, category, tags, trust_score)
|
||||
VALUES (?, ?, ?, ?)
|
||||
""",
|
||||
(
|
||||
content,
|
||||
category,
|
||||
tags,
|
||||
trust_score,
|
||||
canonical_key,
|
||||
metadata_json,
|
||||
confidence if confidence is not None else trust_score,
|
||||
source_role,
|
||||
source_turn,
|
||||
observed_at,
|
||||
contradiction_group,
|
||||
status,
|
||||
),
|
||||
(content, category, tags, self.default_trust),
|
||||
)
|
||||
self._conn.commit()
|
||||
fact_id: int = cur.lastrowid # type: ignore[assignment]
|
||||
except sqlite3.IntegrityError:
|
||||
# Duplicate content — return existing id
|
||||
row = self._conn.execute(
|
||||
"SELECT fact_id FROM facts WHERE content = ?", (content,)
|
||||
).fetchone()
|
||||
return int(row["fact_id"])
|
||||
|
||||
if contradiction_rows:
|
||||
self._mark_contradictions(
|
||||
contradiction_group=contradiction_group,
|
||||
new_canonical_key=canonical_key,
|
||||
existing_rows=contradiction_rows,
|
||||
)
|
||||
|
||||
# Entity extraction and linking
|
||||
for name in self._extract_entities(content):
|
||||
entity_id = self._resolve_entity(name)
|
||||
self._link_fact_entity(fact_id, entity_id)
|
||||
|
||||
# Compute HRR vector after entity linking
|
||||
self._compute_hrr_vector(fact_id, content)
|
||||
self._rebuild_bank(category)
|
||||
|
||||
@@ -342,9 +211,6 @@ class MemoryStore:
|
||||
sql = f"""
|
||||
SELECT f.fact_id, f.content, f.category, f.tags,
|
||||
f.trust_score, f.retrieval_count, f.helpful_count,
|
||||
f.canonical_key, f.metadata_json, f.confidence,
|
||||
f.source_role, f.source_turn, f.observed_at,
|
||||
f.contradiction_group, f.status,
|
||||
f.created_at, f.updated_at
|
||||
FROM facts f
|
||||
JOIN facts_fts fts ON fts.rowid = f.fact_id
|
||||
@@ -470,11 +336,7 @@ class MemoryStore:
|
||||
|
||||
sql = f"""
|
||||
SELECT fact_id, content, category, tags, trust_score,
|
||||
retrieval_count, helpful_count,
|
||||
canonical_key, metadata_json, confidence,
|
||||
source_role, source_turn, observed_at,
|
||||
contradiction_group, status,
|
||||
created_at, updated_at
|
||||
retrieval_count, helpful_count, created_at, updated_at
|
||||
FROM facts
|
||||
WHERE trust_score >= ?
|
||||
{category_clause}
|
||||
@@ -525,89 +387,6 @@ class MemoryStore:
|
||||
"helpful_count": row["helpful_count"] + helpful_increment,
|
||||
}
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# Metadata helpers
|
||||
# ------------------------------------------------------------------
|
||||
|
||||
def _load_metadata(self, metadata_json: str | None) -> dict:
|
||||
if not metadata_json:
|
||||
return {}
|
||||
try:
|
||||
data = json.loads(metadata_json)
|
||||
return data if isinstance(data, dict) else {}
|
||||
except Exception:
|
||||
return {}
|
||||
|
||||
def _merge_metadata(self, existing_json: str | None, incoming: dict | None) -> dict:
|
||||
existing = self._load_metadata(existing_json)
|
||||
incoming = dict(incoming or {})
|
||||
merged = dict(existing)
|
||||
merged.update({k: v for k, v in incoming.items() if k not in {"provenance", "evidence", "observation_count", "duplicate_count", "contradiction_keys"}})
|
||||
|
||||
provenance = []
|
||||
seen_provenance: set[str] = set()
|
||||
for item in list(existing.get("provenance", [])) + list(incoming.get("provenance", [])):
|
||||
if item and item not in seen_provenance:
|
||||
seen_provenance.add(item)
|
||||
provenance.append(item)
|
||||
|
||||
evidence = []
|
||||
seen_evidence: set[tuple[str, str, str]] = set()
|
||||
for item in list(existing.get("evidence", [])) + list(incoming.get("evidence", [])):
|
||||
if not isinstance(item, dict):
|
||||
continue
|
||||
key = (
|
||||
str(item.get("provenance", "")),
|
||||
str(item.get("observed_at", "")),
|
||||
str(item.get("source_text", "")),
|
||||
)
|
||||
if key in seen_evidence:
|
||||
continue
|
||||
seen_evidence.add(key)
|
||||
evidence.append(dict(item))
|
||||
|
||||
observation_count = int(existing.get("observation_count", max(1, len(existing.get("evidence", [])) or 1)))
|
||||
observation_count += int(incoming.get("observation_count", max(1, len(incoming.get("evidence", [])) or 1)))
|
||||
|
||||
contradiction_keys = []
|
||||
seen_keys: set[str] = set()
|
||||
for item in list(existing.get("contradiction_keys", [])) + list(incoming.get("contradiction_keys", [])):
|
||||
if item and item not in seen_keys:
|
||||
seen_keys.add(item)
|
||||
contradiction_keys.append(item)
|
||||
|
||||
merged["provenance"] = provenance
|
||||
merged["evidence"] = evidence
|
||||
merged["observation_count"] = observation_count
|
||||
merged["duplicate_count"] = max(0, observation_count - 1)
|
||||
if contradiction_keys:
|
||||
merged["contradiction_keys"] = contradiction_keys
|
||||
return merged
|
||||
|
||||
def _mark_contradictions(self, contradiction_group: str, new_canonical_key: str, existing_rows: list[sqlite3.Row]) -> None:
|
||||
for row in existing_rows:
|
||||
metadata = self._load_metadata(row["metadata_json"])
|
||||
keys = []
|
||||
seen: set[str] = set()
|
||||
for item in list(metadata.get("contradiction_keys", [])) + [new_canonical_key]:
|
||||
if item and item not in seen:
|
||||
seen.add(item)
|
||||
keys.append(item)
|
||||
metadata["status"] = "contradiction"
|
||||
metadata["contradiction_group"] = contradiction_group
|
||||
metadata["contradiction_keys"] = keys
|
||||
self._conn.execute(
|
||||
"""
|
||||
UPDATE facts
|
||||
SET status = 'contradiction',
|
||||
metadata_json = ?,
|
||||
updated_at = CURRENT_TIMESTAMP
|
||||
WHERE fact_id = ?
|
||||
""",
|
||||
(json.dumps(metadata, sort_keys=True), row["fact_id"]),
|
||||
)
|
||||
self._conn.commit()
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# Entity helpers
|
||||
# ------------------------------------------------------------------
|
||||
@@ -781,14 +560,8 @@ class MemoryStore:
|
||||
# ------------------------------------------------------------------
|
||||
|
||||
def _row_to_dict(self, row: sqlite3.Row) -> dict:
|
||||
"""Convert a sqlite3.Row to a plain dict with decoded metadata."""
|
||||
data = dict(row)
|
||||
metadata = self._load_metadata(data.get("metadata_json"))
|
||||
if metadata:
|
||||
data["metadata"] = metadata
|
||||
data.setdefault("relation", metadata.get("relation"))
|
||||
data.pop("metadata_json", None)
|
||||
return data
|
||||
"""Convert a sqlite3.Row to a plain dict."""
|
||||
return dict(row)
|
||||
|
||||
def close(self) -> None:
|
||||
"""Close the database connection."""
|
||||
|
||||
63
tests/fixtures/memory_extraction_fragments.json
vendored
63
tests/fixtures/memory_extraction_fragments.json
vendored
@@ -1,63 +0,0 @@
|
||||
{
|
||||
"preferences_and_duplicates": [
|
||||
{
|
||||
"role": "user",
|
||||
"content": "Deploy via Ansible for production changes.",
|
||||
"created_at": "2026-04-22T10:00:00Z"
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "We deploy through Ansible on this repo.",
|
||||
"created_at": "2026-04-22T10:01:00Z"
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "Gitea-first for repository work.",
|
||||
"created_at": "2026-04-22T10:02:00Z"
|
||||
}
|
||||
],
|
||||
"operational_and_contradictions": [
|
||||
{
|
||||
"role": "user",
|
||||
"content": "The BURN watchdog caps dispatches per cycle to 6.",
|
||||
"created_at": "2026-04-22T11:00:00Z"
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "The provider should stay openai-codex/gpt-5.4.",
|
||||
"created_at": "2026-04-22T11:01:00Z"
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "Correction: the provider should stay mimo-v2-pro.",
|
||||
"created_at": "2026-04-22T11:02:00Z"
|
||||
}
|
||||
],
|
||||
"mixed_transcript": [
|
||||
{
|
||||
"role": "user",
|
||||
"content": "Deploy via Ansible for production changes.",
|
||||
"created_at": "2026-04-22T10:00:00Z"
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "We deploy through Ansible on this repo.",
|
||||
"created_at": "2026-04-22T10:01:00Z"
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "The BURN watchdog caps dispatches per cycle to 6.",
|
||||
"created_at": "2026-04-22T11:00:00Z"
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "The provider should stay openai-codex/gpt-5.4.",
|
||||
"created_at": "2026-04-22T11:01:00Z"
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "Correction: the provider should stay mimo-v2-pro.",
|
||||
"created_at": "2026-04-22T11:02:00Z"
|
||||
}
|
||||
]
|
||||
}
|
||||
@@ -1,50 +0,0 @@
|
||||
"""Integration tests for holographic auto-extraction with structured fact persistence."""
|
||||
|
||||
import json
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parents[3]))
|
||||
|
||||
from plugins.memory.holographic import HolographicMemoryProvider
|
||||
|
||||
_FIXTURE_PATH = Path(__file__).resolve().parents[2] / "fixtures" / "memory_extraction_fragments.json"
|
||||
|
||||
|
||||
def _load_fixture(name: str):
|
||||
return json.loads(_FIXTURE_PATH.read_text())[name]
|
||||
|
||||
|
||||
class TestHolographicAutoExtract:
|
||||
def test_auto_extract_persists_structured_metadata_and_normalizes_duplicates(self, tmp_path):
|
||||
provider = HolographicMemoryProvider(
|
||||
config={
|
||||
"db_path": str(tmp_path / "memory_store.db"),
|
||||
"auto_extract": True,
|
||||
"default_trust": 0.5,
|
||||
}
|
||||
)
|
||||
provider.initialize("test-session")
|
||||
|
||||
messages = _load_fixture("mixed_transcript")
|
||||
provider.on_session_end(messages)
|
||||
provider.on_session_end(messages)
|
||||
|
||||
facts = provider._store.list_facts(min_trust=0.0, limit=20)
|
||||
deploy_facts = [f for f in facts if f.get("relation") == "workflow.deploy_method"]
|
||||
provider_facts = [f for f in facts if f.get("contradiction_group") == "config.provider"]
|
||||
|
||||
assert len(deploy_facts) == 1
|
||||
assert deploy_facts[0]["metadata"]["duplicate_count"] >= 3
|
||||
assert deploy_facts[0]["observed_at"] == "2026-04-22T10:00:00Z"
|
||||
assert deploy_facts[0]["metadata"]["provenance"] == [
|
||||
"conversation:user:0",
|
||||
"conversation:user:1",
|
||||
]
|
||||
|
||||
assert len(provider_facts) == 2
|
||||
assert {f["status"] for f in provider_facts} == {"contradiction"}
|
||||
assert {f["metadata"]["value"] for f in provider_facts} == {
|
||||
"openai-codex/gpt-5.4",
|
||||
"mimo-v2-pro",
|
||||
}
|
||||
@@ -1,6 +1,6 @@
|
||||
"""Tests for session compaction with fact extraction."""
|
||||
|
||||
import json
|
||||
import pytest
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
@@ -8,19 +8,12 @@ sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
||||
|
||||
from agent.session_compactor import (
|
||||
ExtractedFact,
|
||||
evaluate_extraction_quality,
|
||||
extract_and_save_facts,
|
||||
extract_facts_from_messages,
|
||||
format_facts_summary,
|
||||
save_facts_to_store,
|
||||
extract_and_save_facts,
|
||||
format_facts_summary,
|
||||
)
|
||||
|
||||
_FIXTURE_PATH = Path(__file__).resolve().parent / "fixtures" / "memory_extraction_fragments.json"
|
||||
|
||||
|
||||
def _load_fixture(name: str):
|
||||
return json.loads(_FIXTURE_PATH.read_text())[name]
|
||||
|
||||
|
||||
class TestFactExtraction:
|
||||
def test_extract_preference(self):
|
||||
@@ -67,48 +60,14 @@ class TestFactExtraction:
|
||||
{"role": "user", "content": "I prefer Python."},
|
||||
]
|
||||
facts = extract_facts_from_messages(messages)
|
||||
# Should deduplicate
|
||||
python_facts = [f for f in facts if "Python" in f.content]
|
||||
assert len(python_facts) == 1
|
||||
|
||||
def test_structured_fact_preserves_provenance_and_temporal_metadata(self):
|
||||
facts = extract_facts_from_messages(_load_fixture("preferences_and_duplicates"))
|
||||
deploy_fact = next(f for f in facts if f.relation == "workflow.deploy_method")
|
||||
assert deploy_fact.source_role == "user"
|
||||
assert deploy_fact.source_turn == 0
|
||||
assert deploy_fact.observed_at == "2026-04-22T10:00:00Z"
|
||||
assert deploy_fact.provenance == "conversation:user:0"
|
||||
assert deploy_fact.canonical_key
|
||||
assert deploy_fact.evidence
|
||||
assert deploy_fact.evidence[0]["source_text"].startswith("Deploy via Ansible")
|
||||
|
||||
def test_near_duplicate_facts_are_normalized_into_one_canonical_fact(self):
|
||||
facts = extract_facts_from_messages(_load_fixture("preferences_and_duplicates"))
|
||||
deploy_facts = [f for f in facts if f.relation == "workflow.deploy_method"]
|
||||
assert len(deploy_facts) == 1
|
||||
assert len(deploy_facts[0].evidence) == 2
|
||||
assert deploy_facts[0].metadata["duplicate_count"] == 1
|
||||
|
||||
def test_contradictory_facts_are_preserved_for_unique_slots(self):
|
||||
facts = extract_facts_from_messages(_load_fixture("operational_and_contradictions"))
|
||||
provider_facts = [f for f in facts if f.contradiction_group == "config.provider"]
|
||||
assert len(provider_facts) == 2
|
||||
assert {f.status for f in provider_facts} == {"contradiction"}
|
||||
assert {f.normalized_content for f in provider_facts} == {
|
||||
"openai codex gpt 5 4",
|
||||
"mimo v2 pro",
|
||||
}
|
||||
|
||||
def test_quality_evaluation_reports_noise_reduction(self):
|
||||
metrics = evaluate_extraction_quality(_load_fixture("mixed_transcript"))
|
||||
assert metrics["raw_candidates"] > metrics["normalized_facts"]
|
||||
assert metrics["noise_reduction"] > 0
|
||||
assert metrics["contradiction_groups"] == 1
|
||||
|
||||
|
||||
class TestSaveFacts:
|
||||
def test_save_with_callback(self):
|
||||
saved = []
|
||||
|
||||
def mock_save(category, entity, content, trust):
|
||||
saved.append({"category": category, "content": content})
|
||||
|
||||
@@ -117,38 +76,6 @@ class TestSaveFacts:
|
||||
assert count == 1
|
||||
assert len(saved) == 1
|
||||
|
||||
def test_save_with_extended_callback_metadata(self):
|
||||
saved = []
|
||||
|
||||
def mock_save(category, entity, content, trust, **kwargs):
|
||||
saved.append({
|
||||
"category": category,
|
||||
"entity": entity,
|
||||
"content": content,
|
||||
"trust": trust,
|
||||
**kwargs,
|
||||
})
|
||||
|
||||
fact = ExtractedFact(
|
||||
"project.operational",
|
||||
"watchdog",
|
||||
"BURN watchdog caps dispatches per cycle to 6",
|
||||
0.9,
|
||||
2,
|
||||
source_role="user",
|
||||
observed_at="2026-04-22T11:00:00Z",
|
||||
provenance="conversation:user:2",
|
||||
canonical_key="project.operational|watchdog|dispatch_cap|6",
|
||||
relation="fleet.dispatch_cap",
|
||||
contradiction_group="fleet.dispatch_cap",
|
||||
metadata={"duplicate_count": 0},
|
||||
)
|
||||
count = save_facts_to_store([fact], fact_store_fn=mock_save)
|
||||
assert count == 1
|
||||
assert saved[0]["canonical_key"] == fact.canonical_key
|
||||
assert saved[0]["observed_at"] == "2026-04-22T11:00:00Z"
|
||||
assert saved[0]["metadata"]["duplicate_count"] == 0
|
||||
|
||||
|
||||
class TestFormatSummary:
|
||||
def test_empty(self):
|
||||
|
||||
@@ -199,7 +199,7 @@ class TestMarkdown:
|
||||
class TestDataset:
|
||||
def test_sample_dataset_has_entries(self):
|
||||
dataset = generate_sample_dataset()
|
||||
assert len(dataset) >= 4
|
||||
assert len(dataset) >= 50
|
||||
|
||||
def test_sample_dataset_structure(self):
|
||||
dataset = generate_sample_dataset()
|
||||
@@ -216,6 +216,9 @@ class TestDataset:
|
||||
assert "screenshot" in categories
|
||||
assert "diagram" in categories
|
||||
assert "photo" in categories
|
||||
assert "chart" in categories
|
||||
assert "ocr" in categories
|
||||
assert "document" in categories
|
||||
|
||||
|
||||
class TestModels:
|
||||
|
||||
21
tests/test_vision_benchmark_artifacts.py
Normal file
21
tests/test_vision_benchmark_artifacts.py
Normal file
@@ -0,0 +1,21 @@
|
||||
import json
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
DATASET = Path("benchmarks/test_images.json")
|
||||
REPORT = Path("metrics/vision-benchmark-smoke-2026-04-22.md")
|
||||
|
||||
|
||||
def test_benchmark_dataset_is_issue_sized_and_category_complete() -> None:
|
||||
items = json.loads(DATASET.read_text(encoding="utf-8"))
|
||||
assert len(items) >= 50
|
||||
categories = {item["category"] for item in items}
|
||||
assert {"screenshot", "diagram", "photo", "ocr", "chart", "document"}.issubset(categories)
|
||||
|
||||
|
||||
def test_metrics_report_exists_with_recommendation() -> None:
|
||||
assert REPORT.exists(), "missing benchmark report under metrics/"
|
||||
text = REPORT.read_text(encoding="utf-8")
|
||||
assert "Recommendation" in text
|
||||
assert "Gemma 4" in text
|
||||
assert "Gemini" in text
|
||||
Reference in New Issue
Block a user