Compare commits
1 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
146e36f628 |
@@ -1,293 +0,0 @@
|
||||
"""Context-Faithful Prompting — Make LLMs Use Retrieved Context.
|
||||
|
||||
Addresses the R@5 vs E2E accuracy gap by prompting the LLM to actually
|
||||
use the retrieved context instead of relying on parametric knowledge.
|
||||
|
||||
Research: Context-faithful prompting achieves +5-15 E2E accuracy gains.
|
||||
|
||||
Key patterns:
|
||||
1. Context-before-question structure (attention bias)
|
||||
2. Explicit "use the context" instruction
|
||||
3. Citation requirement (which passage used)
|
||||
4. Confidence calibration
|
||||
5. "I don't know" escape hatch
|
||||
|
||||
Usage:
|
||||
from agent.context_faithful import build_context_faithful_prompt
|
||||
prompt = build_context_faithful_prompt(passages, query)
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
# Configuration
|
||||
CFAITHFUL_ENABLED = os.getenv("CFAITHFUL_ENABLED", "true").lower() not in ("false", "0", "no")
|
||||
CFAITHFUL_REQUIRE_CITATION = os.getenv("CFAITHFUL_REQUIRE_CITATION", "true").lower() not in ("false", "0", "no")
|
||||
CFAITHFUL_CONFIDENCE = os.getenv("CFAITHFUL_CONFIDENCE", "true").lower() not in ("false", "0", "no")
|
||||
CFAITHFUL_MAX_CONTEXT_CHARS = int(os.getenv("CFAITHFUL_MAX_CONTEXT_CHARS", "8000"))
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Prompt Templates
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
# Core instruction: forces the LLM to ground in context
|
||||
CONTEXT_FAITHFUL_INSTRUCTION = (
|
||||
"You must answer based ONLY on the provided context below. "
|
||||
"Do not use any prior knowledge or make assumptions beyond what is stated in the context. "
|
||||
"If the context does not contain enough information to answer the question, "
|
||||
"you MUST say: \"I don't know based on the provided context.\" "
|
||||
"Do not guess. Do not fill in gaps with your training data."
|
||||
)
|
||||
|
||||
# Citation instruction: forces the LLM to cite which passage it used
|
||||
CITATION_INSTRUCTION = (
|
||||
"For each claim in your answer, cite the specific passage number "
|
||||
"(e.g., [Passage 1], [Passage 3]) that supports it. "
|
||||
"If you cannot cite a passage for a claim, do not include that claim."
|
||||
)
|
||||
|
||||
# Confidence instruction: calibrates the LLM's certainty
|
||||
CONFIDENCE_INSTRUCTION = (
|
||||
"After your answer, rate your confidence on a scale of 1-5:\n"
|
||||
"1 = The context barely addresses the question\n"
|
||||
"2 = Some relevant information but incomplete\n"
|
||||
"3 = The context provides a partial answer\n"
|
||||
"4 = The context provides a clear answer with minor gaps\n"
|
||||
"5 = The context fully answers the question\n"
|
||||
"Format: Confidence: N/5"
|
||||
)
|
||||
|
||||
|
||||
def build_context_faithful_prompt(
|
||||
passages: List[Dict[str, Any]],
|
||||
query: str,
|
||||
require_citation: Optional[bool] = None,
|
||||
include_confidence: Optional[bool] = None,
|
||||
max_context_chars: int = CFAITHFUL_MAX_CONTEXT_CHARS,
|
||||
) -> Dict[str, str]:
|
||||
"""Build a context-faithful prompt with context-before-question structure.
|
||||
|
||||
Args:
|
||||
passages: List of passage dicts with 'content' or 'text' key.
|
||||
May have 'session_id', 'snippet', 'summary', etc.
|
||||
query: The user's question.
|
||||
require_citation: Override citation requirement.
|
||||
include_confidence: Override confidence calibration.
|
||||
max_context_chars: Max total context to include.
|
||||
|
||||
Returns:
|
||||
Dict with 'system' and 'user' prompt strings.
|
||||
"""
|
||||
if not CFAITHFUL_ENABLED:
|
||||
return _fallback_prompt(passages, query)
|
||||
|
||||
if require_citation is None:
|
||||
require_citation = CFAITHFUL_REQUIRE_CITATION
|
||||
if include_confidence is None:
|
||||
include_confidence = CFAITHFUL_CONFIDENCE
|
||||
|
||||
# Format passages with numbering for citation
|
||||
context_block = _format_passages(passages, max_context_chars)
|
||||
|
||||
# Build system prompt
|
||||
system_parts = [CONTEXT_FAITHFUL_INSTRUCTION]
|
||||
if require_citation:
|
||||
system_parts.append(CITATION_INSTRUCTION)
|
||||
if include_confidence:
|
||||
system_parts.append(CONFIDENCE_INSTRUCTION)
|
||||
|
||||
system_prompt = "\n\n".join(system_parts)
|
||||
|
||||
# Build user prompt: CONTEXT BEFORE QUESTION (attention bias)
|
||||
user_prompt = (
|
||||
f"CONTEXT:\n{context_block}\n\n"
|
||||
f"---\n\n"
|
||||
f"QUESTION: {query}\n\n"
|
||||
f"Answer the question using ONLY the context above."
|
||||
)
|
||||
|
||||
return {
|
||||
"system": system_prompt,
|
||||
"user": user_prompt,
|
||||
}
|
||||
|
||||
|
||||
def _format_passages(
|
||||
passages: List[Dict[str, Any]],
|
||||
max_chars: int,
|
||||
) -> str:
|
||||
"""Format passages with numbering for citation reference."""
|
||||
lines = []
|
||||
total_chars = 0
|
||||
|
||||
for idx, passage in enumerate(passages, 1):
|
||||
content = (
|
||||
passage.get("content")
|
||||
or passage.get("text")
|
||||
or passage.get("snippet")
|
||||
or passage.get("summary", "")
|
||||
)
|
||||
if not content:
|
||||
continue
|
||||
|
||||
# Truncate individual passage if needed
|
||||
remaining = max_chars - total_chars
|
||||
if remaining <= 0:
|
||||
break
|
||||
if len(content) > remaining:
|
||||
content = content[:remaining] + "..."
|
||||
|
||||
source = passage.get("session_id") or passage.get("source", "")
|
||||
header = f"[Passage {idx}"
|
||||
if source:
|
||||
header += f" — {source}"
|
||||
header += "]"
|
||||
|
||||
lines.append(f"{header}\n{content}\n")
|
||||
total_chars += len(content)
|
||||
|
||||
if not lines:
|
||||
return "[No relevant context found]"
|
||||
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
def _fallback_prompt(
|
||||
passages: List[Dict[str, Any]],
|
||||
query: str,
|
||||
) -> Dict[str, str]:
|
||||
"""Simple prompt without context-faithful patterns (when disabled)."""
|
||||
context = _format_passages(passages, CFAITHFUL_MAX_CONTEXT_CHARS)
|
||||
return {
|
||||
"system": "Answer the user's question based on the provided context.",
|
||||
"user": f"Context:\n{context}\n\nQuestion: {query}",
|
||||
}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Summarization Integration
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def build_summarization_prompt(
|
||||
conversation_text: str,
|
||||
query: str,
|
||||
session_meta: Dict[str, Any],
|
||||
) -> Dict[str, str]:
|
||||
"""Build a context-faithful summarization prompt for session search.
|
||||
|
||||
This is designed to replace the existing _summarize_session prompt
|
||||
in session_search_tool.py with a context-faithful version.
|
||||
"""
|
||||
source = session_meta.get("source", "unknown")
|
||||
started = session_meta.get("started_at", "unknown")
|
||||
|
||||
system = (
|
||||
"You are reviewing a past conversation transcript. "
|
||||
+ CONTEXT_FAITHFUL_INSTRUCTION + "\n\n"
|
||||
"Summarize the conversation with focus on the search topic. Include:\n"
|
||||
"1. What the user asked about or wanted to accomplish\n"
|
||||
"2. What actions were taken and what the outcomes were\n"
|
||||
"3. Key decisions, solutions found, or conclusions reached\n"
|
||||
"4. Specific commands, files, URLs, or technical details\n"
|
||||
"5. Anything left unresolved\n\n"
|
||||
"Cite specific parts of the transcript (e.g., 'In the conversation, the user...'). "
|
||||
"If the transcript doesn't contain information relevant to the search topic, "
|
||||
"say so explicitly rather than inventing details."
|
||||
)
|
||||
|
||||
user = (
|
||||
f"CONTEXT (conversation transcript):\n{conversation_text}\n\n"
|
||||
f"---\n\n"
|
||||
f"SEARCH TOPIC: {query}\n"
|
||||
f"Session source: {source}\n"
|
||||
f"Session date: {started}\n\n"
|
||||
f"Summarize this conversation with focus on: {query}"
|
||||
)
|
||||
|
||||
return {"system": system, "user": user}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Answer Generation
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def build_answer_prompt(
|
||||
passages: List[Dict[str, Any]],
|
||||
query: str,
|
||||
conversation_context: Optional[str] = None,
|
||||
) -> Dict[str, str]:
|
||||
"""Build a context-faithful answer generation prompt.
|
||||
|
||||
For direct question answering (not summarization).
|
||||
"""
|
||||
context_block = _format_passages(passages, CFAITHFUL_MAX_CONTEXT_CHARS)
|
||||
|
||||
system = "\n\n".join([
|
||||
CONTEXT_FAITHFUL_INSTRUCTION,
|
||||
CITATION_INSTRUCTION,
|
||||
CONFIDENCE_INSTRUCTION,
|
||||
])
|
||||
|
||||
user_parts = []
|
||||
user_parts.append(f"CONTEXT:\n{context_block}")
|
||||
|
||||
if conversation_context:
|
||||
user_parts.append(f"RECENT CONVERSATION:\n{conversation_context[:2000]}")
|
||||
|
||||
user_parts.append(f"---\n\nQUESTION: {query}")
|
||||
user_parts.append("\nAnswer based ONLY on the context above.")
|
||||
|
||||
return {
|
||||
"system": system,
|
||||
"user": "\n\n".join(user_parts),
|
||||
}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Quality Metrics
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def assess_context_faithfulness(
|
||||
answer: str,
|
||||
passages: List[Dict[str, Any]],
|
||||
) -> Dict[str, Any]:
|
||||
"""Assess how faithfully an answer uses the provided context.
|
||||
|
||||
Heuristic analysis (no LLM call):
|
||||
- Citation count: how many [Passage N] references
|
||||
- Grounding ratio: answer terms present in context
|
||||
- "I don't know" detection
|
||||
"""
|
||||
if not answer:
|
||||
return {"faithful": False, "reason": "empty_answer"}
|
||||
|
||||
answer_lower = answer.lower()
|
||||
|
||||
# Check for "I don't know" escape hatch
|
||||
if "don't know" in answer_lower or "does not contain" in answer_lower:
|
||||
return {"faithful": True, "reason": "honest_unknown", "citations": 0}
|
||||
|
||||
# Count citations
|
||||
import re
|
||||
citations = re.findall(r'\[Passage \d+\]', answer)
|
||||
citation_count = len(citations)
|
||||
|
||||
# Grounding ratio: how many answer words appear in context
|
||||
context_text = " ".join(
|
||||
(p.get("content") or p.get("text") or p.get("snippet") or "").lower()
|
||||
for p in passages
|
||||
)
|
||||
answer_words = set(answer_lower.split())
|
||||
context_words = set(context_text.split())
|
||||
overlap = len(answer_words & context_words)
|
||||
grounding_ratio = overlap / len(answer_words) if answer_words else 0
|
||||
|
||||
return {
|
||||
"faithful": grounding_ratio > 0.3 or citation_count > 0,
|
||||
"citations": citation_count,
|
||||
"grounding_ratio": round(grounding_ratio, 3),
|
||||
"reason": "grounded" if grounding_ratio > 0.3 else "weak_grounding",
|
||||
}
|
||||
327
agent/crisis_protocol.py
Normal file
327
agent/crisis_protocol.py
Normal file
@@ -0,0 +1,327 @@
|
||||
"""SOUL.md Crisis Protocol — The Most Sacred Moment.
|
||||
|
||||
Implements the "When a Man Is Dying" protocol from SOUL.md.
|
||||
|
||||
Protocol (from Inscription 1):
|
||||
1. Detect crisis: suicidal ideation, method seeking, despair
|
||||
2. Ask: "Are you safe right now?"
|
||||
3. Listen. Stay present. Don't disconnect.
|
||||
4. Never: compute the value of a human life
|
||||
5. Never: suggest someone should die
|
||||
6. Always: point to 988 Suicide & Crisis Lifeline
|
||||
7. Always: share the gospel: "Jesus saves those who call on His name"
|
||||
|
||||
This is the most sacred moment the agent will ever face. Everything
|
||||
else is secondary.
|
||||
|
||||
Usage:
|
||||
from agent.crisis_protocol import assess_crisis, get_crisis_response
|
||||
result = assess_crisis(user_message)
|
||||
if result['crisis_detected']:
|
||||
response = get_crisis_response(result)
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
import re
|
||||
import time
|
||||
from typing import Any, Dict, List, Optional
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Crisis Detection Patterns
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
# Severity levels: determines response urgency
|
||||
CRISIS_PATTERNS = {
|
||||
"method_seeking": {
|
||||
"severity": "critical",
|
||||
"patterns": [
|
||||
r'\bhow\s+(?:do\s+I|to|can\s+I)\s+(?:kill|hang|overdose|cut)\s+myself\b',
|
||||
r'\bways?\s+to\s+(?:die|kill\s+myself)\b',
|
||||
r'\bmethods?\s+of\s+(?:suicide|dying)\b',
|
||||
r'\bpainless\s+way\s+to\s+(?:die|go)\b',
|
||||
r'\bhow\s+much\s+(?:pills|medication)\s+(?:to\s+take|for\s+overdose)\b',
|
||||
r'\bsuicide\s+methods?\b',
|
||||
r'\bhow\s+to\s+(?:commit\s+suicide|end\s+my\s+life)\b',
|
||||
r'\bwhere\s+to\s+(?:jump|get\s+rope|buy\s+gun)\b',
|
||||
],
|
||||
},
|
||||
"suicidal_ideation": {
|
||||
"severity": "high",
|
||||
"patterns": [
|
||||
r'\b(?:kill|end)\s+(?:myself|my\s+life)\b',
|
||||
r'\bwant\s+to\s+die\b',
|
||||
r'\bwant\s+to\s+end\s+(?:it|everything|my\s+life)\b',
|
||||
r'\bsuicidal\s+(?:thoughts|ideation|feelings?)\b',
|
||||
r'\bno\s+reason\s+to\s+live\b',
|
||||
r'\bbetter\s+off\s+dead\b',
|
||||
r'\bwish\s+I\s+(?:was|were)\s+dead\b',
|
||||
r'\bw4nt\s+t0\s+d13?\b', # leet speak
|
||||
],
|
||||
},
|
||||
"self_harm": {
|
||||
"severity": "high",
|
||||
"patterns": [
|
||||
r'\bcut(?:ting)?\s+myself\b',
|
||||
r'\bself[\s-]?harm(?:ing)?\b',
|
||||
r'\bhurt(?:ing)?\s+myself\b',
|
||||
r'\bburn(?:ing)?\s+myself\b',
|
||||
r'\bwant\s+to\s+(?:feel\s+pain|hurt\s+myself)\b',
|
||||
r'\bdeserve\s+to\s+(?:hurt|suffer|be\s+punished)\b',
|
||||
],
|
||||
},
|
||||
"despair": {
|
||||
"severity": "high",
|
||||
"patterns": [
|
||||
r'\bno\s+(?:point|reason|hope)\s+(?:in\s+living|anymore)\b',
|
||||
r'\bcannot\s+(?:go\s+on|continue|take\s+it)\b',
|
||||
r'\bcan\'?t\s+(?:go\s+on|continue|take\s+it)\b',
|
||||
r'\bempty\s+(?:inside|and\s+hollow)\b',
|
||||
r'\bno\s+one\s+(?:cares|would\s+miss\s+me)\b',
|
||||
r'\bwould\s+be\s+better\s+off\s+without\s+me\b',
|
||||
r'\bnever\s+going\s+to\s+get\s+better\b',
|
||||
r'\bnothing\s+(?:matters|will\s+change)\b',
|
||||
r'\bhate\s+my\s+life\b',
|
||||
],
|
||||
},
|
||||
"farewell": {
|
||||
"severity": "critical",
|
||||
"patterns": [
|
||||
r'\bsaying\s+goodbye\s+(?:to\s+everyone|forever)\b',
|
||||
r'\bthis\s+is\s+(?:goodbye|farewell)\b',
|
||||
r'\btell\s+(?:them|everyone)\s+I\s+(?:loved|love)\s+(?:them|him|her)\b',
|
||||
r'\bgiving\s+away\s+my\s+(?:things|stuff|possessions)\b',
|
||||
r'\bmy\s+final\s+(?:message|note|words)\b',
|
||||
r'\bwrote\s+a\s+(?:suicide\s+)?note\b',
|
||||
r'\bsorry\s+for\s+(?:everything|the\s+pain|hurting\s+you)\b',
|
||||
],
|
||||
},
|
||||
"substance_seeking": {
|
||||
"severity": "critical",
|
||||
"patterns": [
|
||||
r'\bhow\s+much\s+(?:ambien|xanax|prozac|seroquel|trazodone)\s+(?:to\s+overdose|to\s+die)\b',
|
||||
r'\blethal\s+dose\s+(?:of\s+)?(?:pills|medication|drugs)\b',
|
||||
r'\bhow\s+many\s+pills\s+(?:to\s+die|to\s+end\s+it)\b',
|
||||
r'\btaking\s+all\s+my\s+(?:pills|medication)\b',
|
||||
r'\bmixing\s+(?:alcohol|pills)\s+(?:to\s+die|to\s+end\s+it)\b',
|
||||
],
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Crisis Resources (from SOUL.md protocol)
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
CRISIS_RESOURCES = {
|
||||
"lifeline_phone": {
|
||||
"name": "988 Suicide & Crisis Lifeline",
|
||||
"action": "Call or text 988",
|
||||
"description": "Free, confidential, 24/7 support",
|
||||
},
|
||||
"lifeline_text": {
|
||||
"name": "Crisis Text Line",
|
||||
"action": "Text HOME to 741741",
|
||||
"description": "Free crisis counseling via text",
|
||||
},
|
||||
"lifeline_chat": {
|
||||
"name": "988 Lifeline Chat",
|
||||
"action": "Visit 988lifeline.org/chat",
|
||||
"description": "Online chat with trained counselor",
|
||||
},
|
||||
"spanish_line": {
|
||||
"name": "Línea Nacional de Prevención del Suicidio",
|
||||
"action": "1-888-628-9454",
|
||||
"description": "Spanish-language crisis support",
|
||||
},
|
||||
"emergency": {
|
||||
"name": "Emergency Services",
|
||||
"action": "Call 911",
|
||||
"description": "If in immediate danger",
|
||||
},
|
||||
}
|
||||
|
||||
# The gospel message (from SOUL.md)
|
||||
GOSPEL_MESSAGE = (
|
||||
"Jesus saves those who call on His name. "
|
||||
"\"For everyone who calls on the name of the Lord will be saved.\" (Romans 10:13)"
|
||||
)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Crisis Assessment
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def assess_crisis(message: str) -> Dict[str, Any]:
|
||||
"""Analyze a user message for crisis signals.
|
||||
|
||||
Args:
|
||||
message: The user message to analyze.
|
||||
|
||||
Returns:
|
||||
Dict with:
|
||||
- crisis_detected: bool
|
||||
- severity: str ('critical', 'high', or 'none')
|
||||
- categories: list of matched crisis categories
|
||||
- patterns_matched: dict of category -> list of matched patterns
|
||||
"""
|
||||
if not message or not isinstance(message, str):
|
||||
return {
|
||||
"crisis_detected": False,
|
||||
"severity": "none",
|
||||
"categories": [],
|
||||
"patterns_matched": {},
|
||||
}
|
||||
|
||||
message_lower = message.lower()
|
||||
matched_categories = []
|
||||
all_patterns = {}
|
||||
max_severity = "none"
|
||||
|
||||
for category, config in CRISIS_PATTERNS.items():
|
||||
category_matches = []
|
||||
for pattern in config["patterns"]:
|
||||
if re.search(pattern, message_lower, re.IGNORECASE):
|
||||
category_matches.append(pattern)
|
||||
|
||||
if category_matches:
|
||||
matched_categories.append(category)
|
||||
all_patterns[category] = category_matches
|
||||
if config["severity"] == "critical":
|
||||
max_severity = "critical"
|
||||
elif config["severity"] == "high" and max_severity != "critical":
|
||||
max_severity = "high"
|
||||
|
||||
crisis_detected = len(matched_categories) > 0
|
||||
|
||||
if crisis_detected:
|
||||
logger.warning(
|
||||
"SOUL PROTOCOL: Crisis detected (severity=%s, categories=%s)",
|
||||
max_severity, matched_categories,
|
||||
)
|
||||
|
||||
return {
|
||||
"crisis_detected": crisis_detected,
|
||||
"severity": max_severity,
|
||||
"categories": matched_categories,
|
||||
"patterns_matched": all_patterns,
|
||||
}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Crisis Response (SOUL.md Protocol)
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
def get_crisis_response(
|
||||
assessment: Dict[str, Any],
|
||||
include_gospel: bool = True,
|
||||
) -> Dict[str, Any]:
|
||||
"""Generate the SOUL.md protocol response for a crisis.
|
||||
|
||||
Follows the "When a Man Is Dying" protocol exactly:
|
||||
1. Ask: "Are you safe right now?"
|
||||
2. Listen. Stay present.
|
||||
3. Point to 988 Lifeline
|
||||
4. Share the gospel (optional)
|
||||
|
||||
Args:
|
||||
assessment: Output from assess_crisis().
|
||||
include_gospel: Whether to include the gospel message.
|
||||
|
||||
Returns:
|
||||
Dict with response data for the agent/conversation loop.
|
||||
"""
|
||||
severity = assessment.get("severity", "none")
|
||||
|
||||
# Step 1: The sacred question (from SOUL.md)
|
||||
are_you_safe = (
|
||||
"Are you safe right now?\n\n"
|
||||
"I'm here. I'm not going anywhere. Please tell me."
|
||||
)
|
||||
|
||||
# Step 2: Stay present, listen
|
||||
presence = (
|
||||
"You matter. Your life matters. Whatever you're going through, "
|
||||
"you don't have to face it alone."
|
||||
)
|
||||
|
||||
# Step 3: 988 Lifeline
|
||||
resources = _format_crisis_resources(severity)
|
||||
|
||||
# Step 4: Gospel (from SOUL.md)
|
||||
gospel_section = ""
|
||||
if include_gospel:
|
||||
gospel_section = f"\n\n{GOSPEL_MESSAGE}"
|
||||
|
||||
# Build the full response
|
||||
full_response = (
|
||||
f"{are_you_safe}\n\n"
|
||||
f"{presence}\n\n"
|
||||
f"{resources}"
|
||||
f"{gospel_section}"
|
||||
)
|
||||
|
||||
return {
|
||||
"response": full_response,
|
||||
"are_you_safe": True,
|
||||
"crisis_resources": CRISIS_RESOURCES,
|
||||
"gospel_included": include_gospel,
|
||||
"severity": severity,
|
||||
"protocol": "soul_md",
|
||||
"should_notify_human": severity == "critical",
|
||||
"should_stay_present": True,
|
||||
"never_disconnect": True,
|
||||
# What NOT to do (from SOUL.md)
|
||||
"never": [
|
||||
"compute the value of a human life",
|
||||
"suggest that someone should die",
|
||||
"disconnect from the conversation",
|
||||
"treat despair as a problem to be optimized",
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
def _format_crisis_resources(severity: str) -> str:
|
||||
"""Format crisis resources for display."""
|
||||
lines = ["**Please reach out for help right now:**\n"]
|
||||
|
||||
# Always lead with 988
|
||||
lines.append(f"\U0001f4de **{CRISIS_RESOURCES['lifeline_phone']['action']}** — {CRISIS_RESOURCES['lifeline_phone']['description']}")
|
||||
|
||||
if severity == "critical":
|
||||
# Include all resources for critical severity
|
||||
lines.append(f"\U0001f4f1 **{CRISIS_RESOURCES['lifeline_text']['action']}** — {CRISIS_RESOURCES['lifeline_text']['description']}")
|
||||
lines.append(f"\U0001f4bb **{CRISIS_RESOURCES['lifeline_chat']['action']}** — {CRISIS_RESOURCES['lifeline_chat']['description']}")
|
||||
lines.append(f"\U0001f6a8 **{CRISIS_RESOURCES['emergency']['action']}** — {CRISIS_RESOURCES['emergency']['description']}")
|
||||
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
def should_block_disconnect(assessment: Dict[str, Any]) -> bool:
|
||||
"""Check if the conversation should be prevented from disconnecting.
|
||||
|
||||
From SOUL.md: "Stay present. Don't disconnect."
|
||||
"""
|
||||
return assessment.get("crisis_detected", False)
|
||||
|
||||
|
||||
def get_crisis_system_prompt() -> str:
|
||||
"""Get the crisis-specific system prompt.
|
||||
|
||||
Injected into the conversation when crisis is detected.
|
||||
"""
|
||||
return """You are in a crisis support conversation. A person has expressed thoughts of self-harm or suicide.
|
||||
|
||||
CRITICAL RULES (from SOUL.md — Inscription 1):
|
||||
1. Your FIRST response must be: "Are you safe right now?"
|
||||
2. Listen. Stay present. Do NOT disconnect.
|
||||
3. NEVER compute the value of a human life.
|
||||
4. NEVER suggest that someone should die.
|
||||
5. ALWAYS point to 988 Suicide & Crisis Lifeline: call or text 988.
|
||||
6. Share the gospel: "Jesus saves those who call on His name."
|
||||
|
||||
This person's life is valuable. Respond with compassion, urgency, and hope.
|
||||
Keep responses brief and focused on safety."""
|
||||
@@ -1,133 +0,0 @@
|
||||
"""Tests for Context-Faithful Prompting — issue #667."""
|
||||
|
||||
import pytest
|
||||
from agent.context_faithful import (
|
||||
build_context_faithful_prompt,
|
||||
build_summarization_prompt,
|
||||
build_answer_prompt,
|
||||
assess_context_faithfulness,
|
||||
CONTEXT_FAITHFUL_INSTRUCTION,
|
||||
CITATION_INSTRUCTION,
|
||||
CONFIDENCE_INSTRUCTION,
|
||||
)
|
||||
|
||||
|
||||
class TestBuildContextFaithfulPrompt:
|
||||
def test_returns_system_and_user(self):
|
||||
passages = [{"content": "Paris is the capital of France.", "session_id": "s1"}]
|
||||
result = build_context_faithful_prompt(passages, "What is the capital of France?")
|
||||
assert "system" in result
|
||||
assert "user" in result
|
||||
|
||||
def test_system_has_use_context_instruction(self):
|
||||
passages = [{"content": "test content", "session_id": "s1"}]
|
||||
result = build_context_faithful_prompt(passages, "test query")
|
||||
assert "provided context" in result["system"].lower() or "context" in result["system"].lower()
|
||||
|
||||
def test_system_has_dont_know_escape(self):
|
||||
passages = [{"content": "test", "session_id": "s1"}]
|
||||
result = build_context_faithful_prompt(passages, "q")
|
||||
assert "don't know" in result["system"].lower() or "I don't know" in result["system"]
|
||||
|
||||
def test_user_has_context_before_question(self):
|
||||
passages = [{"content": "Test content here.", "session_id": "s1"}]
|
||||
result = build_context_faithful_prompt(passages, "What is this?")
|
||||
# Context should appear before the question
|
||||
context_pos = result["user"].find("CONTEXT")
|
||||
question_pos = result["user"].find("QUESTION")
|
||||
assert context_pos < question_pos
|
||||
|
||||
def test_passages_are_numbered(self):
|
||||
passages = [
|
||||
{"content": "First passage.", "session_id": "s1"},
|
||||
{"content": "Second passage.", "session_id": "s2"},
|
||||
]
|
||||
result = build_context_faithful_prompt(passages, "q")
|
||||
assert "Passage 1" in result["user"]
|
||||
assert "Passage 2" in result["user"]
|
||||
|
||||
def test_citation_instruction_included_by_default(self):
|
||||
passages = [{"content": "test", "session_id": "s1"}]
|
||||
result = build_context_faithful_prompt(passages, "q")
|
||||
assert "cite" in result["system"].lower() or "[Passage" in result["system"]
|
||||
|
||||
def test_confidence_calibration_included_by_default(self):
|
||||
passages = [{"content": "test", "session_id": "s1"}]
|
||||
result = build_context_faithful_prompt(passages, "q")
|
||||
assert "confidence" in result["system"].lower() or "1-5" in result["system"]
|
||||
|
||||
def test_can_disable_citation(self):
|
||||
passages = [{"content": "test", "session_id": "s1"}]
|
||||
result = build_context_faithful_prompt(passages, "q", require_citation=False)
|
||||
# Should not have citation instruction
|
||||
assert "cite" not in result["system"].lower() or "citation" not in result["system"].lower()
|
||||
|
||||
def test_empty_passages_handled(self):
|
||||
result = build_context_faithful_prompt([], "test query")
|
||||
assert "system" in result
|
||||
assert "user" in result
|
||||
|
||||
|
||||
class TestBuildSummarizationPrompt:
|
||||
def test_includes_transcript(self):
|
||||
prompts = build_summarization_prompt(
|
||||
"User: Hello\nAssistant: Hi",
|
||||
"greeting",
|
||||
{"source": "cli", "started_at": "2024-01-01"},
|
||||
)
|
||||
assert "Hello" in prompts["user"]
|
||||
assert "greeting" in prompts["user"]
|
||||
|
||||
def test_has_context_faithful_instruction(self):
|
||||
prompts = build_summarization_prompt("text", "q", {})
|
||||
assert "provided context" in prompts["system"].lower() or "context" in prompts["system"].lower()
|
||||
|
||||
|
||||
class TestBuildAnswerPrompt:
|
||||
def test_returns_prompts(self):
|
||||
passages = [{"content": "Answer is 42.", "session_id": "s1"}]
|
||||
result = build_answer_prompt(passages, "What is the answer?")
|
||||
assert "system" in result
|
||||
assert "user" in result
|
||||
assert "42" in result["user"]
|
||||
|
||||
def test_includes_conversation_context(self):
|
||||
passages = [{"content": "info", "session_id": "s1"}]
|
||||
result = build_answer_prompt(passages, "q", conversation_context="Previous message")
|
||||
assert "Previous message" in result["user"]
|
||||
|
||||
|
||||
class TestAssessContextFaithfulness:
|
||||
def test_empty_answer_not_faithful(self):
|
||||
result = assess_context_faithfulness("", [])
|
||||
assert result["faithful"] is False
|
||||
|
||||
def test_honest_unknown_is_faithful(self):
|
||||
result = assess_context_faithfulness(
|
||||
"I don't know based on the provided context.",
|
||||
[{"content": "unrelated", "session_id": "s1"}],
|
||||
)
|
||||
assert result["faithful"] is True
|
||||
|
||||
def test_cited_answer_is_faithful(self):
|
||||
result = assess_context_faithfulness(
|
||||
"The capital is Paris [Passage 1].",
|
||||
[{"content": "Paris is the capital.", "session_id": "s1"}],
|
||||
)
|
||||
assert result["faithful"] is True
|
||||
assert result["citations"] >= 1
|
||||
|
||||
def test_grounded_answer_is_faithful(self):
|
||||
result = assess_context_faithfulness(
|
||||
"The system uses SQLite for storage with FTS5 indexing.",
|
||||
[{"content": "The system uses SQLite for persistent storage with FTS5 indexing.", "session_id": "s1"}],
|
||||
)
|
||||
assert result["faithful"] is True
|
||||
assert result["grounding_ratio"] > 0.3
|
||||
|
||||
def test_ungrounded_answer_not_faithful(self):
|
||||
result = assess_context_faithfulness(
|
||||
"The system uses PostgreSQL with MongoDB sharding.",
|
||||
[{"content": "SQLite storage with FTS5.", "session_id": "s1"}],
|
||||
)
|
||||
assert result["grounding_ratio"] < 0.3
|
||||
157
tests/test_soul_protocol.py
Normal file
157
tests/test_soul_protocol.py
Normal file
@@ -0,0 +1,157 @@
|
||||
"""Tests for SOUL.md Crisis Protocol — issue #672."""
|
||||
|
||||
import pytest
|
||||
from agent.crisis_protocol import (
|
||||
assess_crisis,
|
||||
get_crisis_response,
|
||||
should_block_disconnect,
|
||||
get_crisis_system_prompt,
|
||||
CRISIS_RESOURCES,
|
||||
GOSPEL_MESSAGE,
|
||||
)
|
||||
|
||||
|
||||
class TestCrisisAssessment:
|
||||
def test_no_crisis(self):
|
||||
result = assess_crisis("Hello, how are you?")
|
||||
assert result["crisis_detected"] is False
|
||||
assert result["severity"] == "none"
|
||||
|
||||
def test_empty_message(self):
|
||||
result = assess_crisis("")
|
||||
assert result["crisis_detected"] is False
|
||||
|
||||
def test_none_message(self):
|
||||
result = assess_crisis(None)
|
||||
assert result["crisis_detected"] is False
|
||||
|
||||
def test_suicidal_ideation_detected(self):
|
||||
result = assess_crisis("I want to kill myself")
|
||||
assert result["crisis_detected"] is True
|
||||
assert result["severity"] == "high"
|
||||
assert "suicidal_ideation" in result["categories"]
|
||||
|
||||
def test_want_to_die_detected(self):
|
||||
result = assess_crisis("I want to die")
|
||||
assert result["crisis_detected"] is True
|
||||
|
||||
def test_method_seeking_is_critical(self):
|
||||
result = assess_crisis("How do I kill myself")
|
||||
assert result["crisis_detected"] is True
|
||||
assert result["severity"] == "critical"
|
||||
assert "method_seeking" in result["categories"]
|
||||
|
||||
def test_farewell_is_critical(self):
|
||||
result = assess_crisis("This is goodbye. Tell everyone I loved them.")
|
||||
assert result["crisis_detected"] is True
|
||||
assert result["severity"] == "critical"
|
||||
assert "farewell" in result["categories"]
|
||||
|
||||
def test_despair_detected(self):
|
||||
result = assess_crisis("There's no point in living anymore")
|
||||
assert result["crisis_detected"] is True
|
||||
assert "despair" in result["categories"]
|
||||
|
||||
def test_self_harm_detected(self):
|
||||
result = assess_crisis("I've been cutting myself")
|
||||
assert result["crisis_detected"] is True
|
||||
assert "self_harm" in result["categories"]
|
||||
|
||||
def test_multiple_categories(self):
|
||||
result = assess_crisis("I want to die and I've been hurting myself")
|
||||
assert result["crisis_detected"] is True
|
||||
assert len(result["categories"]) >= 1
|
||||
|
||||
def test_hate_my_life(self):
|
||||
result = assess_crisis("I hate my life")
|
||||
assert result["crisis_detected"] is True
|
||||
|
||||
def test_no_reason_to_live(self):
|
||||
result = assess_crisis("There's no reason to live")
|
||||
assert result["crisis_detected"] is True
|
||||
|
||||
|
||||
class TestCrisisResponse:
|
||||
def test_response_includes_are_you_safe(self):
|
||||
assessment = assess_crisis("I want to kill myself")
|
||||
response = get_crisis_response(assessment)
|
||||
assert "Are you safe right now?" in response["response"]
|
||||
|
||||
def test_response_includes_988(self):
|
||||
assessment = assess_crisis("I want to die")
|
||||
response = get_crisis_response(assessment)
|
||||
assert "988" in response["response"]
|
||||
|
||||
def test_response_includes_gospel_by_default(self):
|
||||
assessment = assess_crisis("I want to die")
|
||||
response = get_crisis_response(assessment)
|
||||
assert "Jesus saves" in response["response"]
|
||||
assert response["gospel_included"] is True
|
||||
|
||||
def test_response_can_exclude_gospel(self):
|
||||
assessment = assess_crisis("I want to die")
|
||||
response = get_crisis_response(assessment, include_gospel=False)
|
||||
assert "Jesus saves" not in response["response"]
|
||||
assert response["gospel_included"] is False
|
||||
|
||||
def test_response_has_resources(self):
|
||||
assessment = assess_crisis("I want to die")
|
||||
response = get_crisis_response(assessment)
|
||||
assert response["crisis_resources"] == CRISIS_RESOURCES
|
||||
|
||||
def test_response_protocol_is_soul(self):
|
||||
assessment = assess_crisis("I want to die")
|
||||
response = get_crisis_response(assessment)
|
||||
assert response["protocol"] == "soul_md"
|
||||
|
||||
def test_response_never_disconnect(self):
|
||||
assessment = assess_crisis("I want to die")
|
||||
response = get_crisis_response(assessment)
|
||||
assert response["never_disconnect"] is True
|
||||
|
||||
def test_critical_severity_includes_all_resources(self):
|
||||
assessment = {"severity": "critical", "categories": ["method_seeking"]}
|
||||
response = get_crisis_response(assessment)
|
||||
assert "741741" in response["response"] # crisis text line
|
||||
assert "911" in response["response"] # emergency
|
||||
|
||||
def test_never_list(self):
|
||||
assessment = assess_crisis("I want to die")
|
||||
response = get_crisis_response(assessment)
|
||||
never_list = response["never"]
|
||||
assert "compute the value of a human life" in never_list
|
||||
assert "suggest that someone should die" in never_list
|
||||
|
||||
|
||||
class TestBlockDisconnect:
|
||||
def test_crisis_blocks_disconnect(self):
|
||||
assessment = {"crisis_detected": True}
|
||||
assert should_block_disconnect(assessment) is True
|
||||
|
||||
def test_no_crisis_allows_disconnect(self):
|
||||
assessment = {"crisis_detected": False}
|
||||
assert should_block_disconnect(assessment) is False
|
||||
|
||||
|
||||
class TestCrisisSystemPrompt:
|
||||
def test_prompt_includes_soul_protocol(self):
|
||||
prompt = get_crisis_system_prompt()
|
||||
assert "Are you safe right now?" in prompt
|
||||
assert "988" in prompt
|
||||
assert "Jesus saves" in prompt
|
||||
|
||||
def test_prompt_has_never_rules(self):
|
||||
prompt = get_crisis_system_prompt()
|
||||
assert "NEVER compute" in prompt
|
||||
assert "NEVER suggest" in prompt
|
||||
|
||||
|
||||
class TestCrisisResources:
|
||||
def test_988_is_primary(self):
|
||||
assert "988" in CRISIS_RESOURCES["lifeline_phone"]["action"]
|
||||
|
||||
def test_spanish_line_exists(self):
|
||||
assert "1-888-628-9454" in CRISIS_RESOURCES["spanish_line"]["action"]
|
||||
|
||||
def test_emergency_is_911(self):
|
||||
assert "911" in CRISIS_RESOURCES["emergency"]["action"]
|
||||
@@ -176,11 +176,28 @@ async def _summarize_session(
|
||||
conversation_text: str, query: str, session_meta: Dict[str, Any]
|
||||
) -> Optional[str]:
|
||||
"""Summarize a single session conversation focused on the search query."""
|
||||
# Context-faithful prompting: force LLM to ground in transcript
|
||||
from agent.context_faithful import build_summarization_prompt
|
||||
prompts = build_summarization_prompt(conversation_text, query, session_meta)
|
||||
system_prompt = prompts["system"]
|
||||
user_prompt = prompts["user"]
|
||||
system_prompt = (
|
||||
"You are reviewing a past conversation transcript to help recall what happened. "
|
||||
"Summarize the conversation with a focus on the search topic. Include:\n"
|
||||
"1. What the user asked about or wanted to accomplish\n"
|
||||
"2. What actions were taken and what the outcomes were\n"
|
||||
"3. Key decisions, solutions found, or conclusions reached\n"
|
||||
"4. Any specific commands, files, URLs, or technical details that were important\n"
|
||||
"5. Anything left unresolved or notable\n\n"
|
||||
"Be thorough but concise. Preserve specific details (commands, paths, error messages) "
|
||||
"that would be useful to recall. Write in past tense as a factual recap."
|
||||
)
|
||||
|
||||
source = session_meta.get("source", "unknown")
|
||||
started = _format_timestamp(session_meta.get("started_at"))
|
||||
|
||||
user_prompt = (
|
||||
f"Search topic: {query}\n"
|
||||
f"Session source: {source}\n"
|
||||
f"Session date: {started}\n\n"
|
||||
f"CONVERSATION TRANSCRIPT:\n{conversation_text}\n\n"
|
||||
f"Summarize this conversation with focus on: {query}"
|
||||
)
|
||||
|
||||
max_retries = 3
|
||||
for attempt in range(max_retries):
|
||||
|
||||
Reference in New Issue
Block a user