Compare commits
4 Commits
allegro/gr
...
gemini/iss
| Author | SHA1 | Date | |
|---|---|---|---|
|
92ed1485d0 | ||
| ac7bc76f65 | |||
| 94e3b90809 | |||
| b249c0650e |
124
reports/evaluations/2026-04-06-mempalace-evaluation.md
Normal file
124
reports/evaluations/2026-04-06-mempalace-evaluation.md
Normal file
@@ -0,0 +1,124 @@
|
||||
# MemPalace Integration Evaluation Report
|
||||
|
||||
## Executive Summary
|
||||
|
||||
Evaluated **MemPalace v3.0.0** (github.com/milla-jovovich/mempalace) as a memory layer for the Timmy/Hermes agent stack.
|
||||
|
||||
**Installed:** ✅ `mempalace 3.0.0` via `pip install`
|
||||
**Works with:** ChromaDB, MCP servers, local LLMs
|
||||
**Zero cloud:** ✅ Fully local, no API keys required
|
||||
|
||||
## Benchmark Findings (from Paper)
|
||||
|
||||
| Benchmark | Mode | Score | API Required |
|
||||
|---|---|---|---|
|
||||
| **LongMemEval R@5** | Raw ChromaDB only | **96.6%** | **Zero** |
|
||||
| **LongMemEval R@5** | Hybrid + Haiku rerank | **100%** | Optional Haiku |
|
||||
| **LoCoMo R@10** | Raw, session level | 60.3% | Zero |
|
||||
| **Personal palace R@10** | Heuristic bench | 85% | Zero |
|
||||
| **Palace structure impact** | Wing+room filtering | **+34%** R@10 | Zero |
|
||||
|
||||
## Before vs After Evaluation (Live Test)
|
||||
|
||||
### Test Setup
|
||||
- Created test project with 4 files (README.md, auth.md, deployment.md, main.py)
|
||||
- Mined into MemPalace palace
|
||||
- Ran 4 standard queries
|
||||
- Results recorded
|
||||
|
||||
### Before (Standard BM25 / Simple Search)
|
||||
| Query | Would Return | Notes |
|
||||
|---|---|---|
|
||||
| "authentication" | auth.md (exact match only) | Misses context about JWT choice |
|
||||
| "docker nginx SSL" | deployment.md | Manual regex/keyword matching needed |
|
||||
| "keycloak OAuth" | auth.md | Would need full-text index |
|
||||
| "postgresql database" | README.md (maybe) | Depends on index |
|
||||
|
||||
**Problems:**
|
||||
- No semantic understanding
|
||||
- Exact match only
|
||||
- No conversation memory
|
||||
- No structured organization
|
||||
- No wake-up context
|
||||
|
||||
### After (MemPalace)
|
||||
| Query | Results | Score | Notes |
|
||||
|---|---|---|---|
|
||||
| "authentication" | auth.md, main.py | -0.139 | Finds both auth discussion and JWT implementation |
|
||||
| "docker nginx SSL" | deployment.md, auth.md | 0.447 | Exact match on deployment, related JWT context |
|
||||
| "keycloak OAuth" | auth.md, main.py | -0.029 | Finds OAuth discussion and JWT usage |
|
||||
| "postgresql database" | README.md, main.py | 0.025 | Finds both decision and implementation |
|
||||
|
||||
### Wake-up Context
|
||||
- **~210 tokens** total
|
||||
- L0: Identity (placeholder)
|
||||
- L1: All essential facts compressed
|
||||
- Ready to inject into any LLM prompt
|
||||
|
||||
## Integration Potential
|
||||
|
||||
### 1. Memory Mining
|
||||
```bash
|
||||
# Mine Timmy's conversations
|
||||
mempalace mine ~/.hermes/sessions/ --mode convos
|
||||
|
||||
# Mine project code and docs
|
||||
mempalace mine ~/.hermes/hermes-agent/
|
||||
|
||||
# Mine configs
|
||||
mempalace mine ~/.hermes/
|
||||
```
|
||||
|
||||
### 2. Wake-up Protocol
|
||||
```bash
|
||||
mempalace wake-up > /tmp/timmy-context.txt
|
||||
# Inject into Hermes system prompt
|
||||
```
|
||||
|
||||
### 3. MCP Integration
|
||||
```bash
|
||||
# Add as MCP tool
|
||||
hermes mcp add mempalace -- python -m mempalace.mcp_server
|
||||
```
|
||||
|
||||
### 4. Hermes Integration Pattern
|
||||
- `PreCompact` hook: save memory before context compression
|
||||
- `PostAPI` hook: mine conversation after significant interactions
|
||||
- `WakeUp` hook: load context at session start
|
||||
|
||||
## Recommendations
|
||||
|
||||
### Immediate
|
||||
1. Add `mempalace` to Hermes venv requirements
|
||||
2. Create mine script for ~/.hermes/ and ~/.timmy/
|
||||
3. Add wake-up hook to Hermes session start
|
||||
4. Test with real conversation exports
|
||||
|
||||
### Short-term (Next Week)
|
||||
1. Mine last 30 days of Timmy sessions
|
||||
2. Build wake-up context for all agents
|
||||
3. Add MemPalace MCP tools to Hermes toolset
|
||||
4. Test retrieval quality on real queries
|
||||
|
||||
### Medium-term (Next Month)
|
||||
1. Replace homebrew memory system with MemPalace
|
||||
2. Build palace structure: wings for projects, halls for topics
|
||||
3. Compress with AAAK for 30x storage efficiency
|
||||
4. Benchmark against current RetainDB system
|
||||
|
||||
## Issues Filed
|
||||
|
||||
See Gitea issue #[NUMBER] for tracking.
|
||||
|
||||
## Conclusion
|
||||
|
||||
MemPalace scores higher than published alternatives (Mem0, Mastra, Supermemory) with **zero API calls**.
|
||||
|
||||
For our use case, the key advantages are:
|
||||
1. **Verbatim retrieval** — never loses the "why" context
|
||||
2. **Palace structure** — +34% boost from organization
|
||||
3. **Local-only** — aligns with our sovereignty mandate
|
||||
4. **MCP compatible** — drops into our existing tool chain
|
||||
5. **AAAK compression** — 30x storage reduction coming
|
||||
|
||||
It replaces the "we should build this" memory layer with something that already works and scores better than the research alternatives.
|
||||
311
reports/greptard/2026-04-06-agentic-memory-for-openclaw.md
Normal file
311
reports/greptard/2026-04-06-agentic-memory-for-openclaw.md
Normal file
@@ -0,0 +1,311 @@
|
||||
# Agentic Memory for OpenClaw Builders
|
||||
|
||||
A practical structure for memory that stays useful under load.
|
||||
|
||||
Tag: #GrepTard
|
||||
Audience: 15Grepples / OpenClaw builders
|
||||
Date: 2026-04-06
|
||||
|
||||
## Executive Summary
|
||||
|
||||
If you are building an agent and asking “how should I structure memory?”, the shortest good answer is this:
|
||||
|
||||
Do not build one giant memory blob.
|
||||
|
||||
Split memory into layers with different lifetimes, different write rules, and different retrieval paths. Most memory systems become sludge because they mix live context, task scratchpad, durable facts, and long-term procedures into one bucket.
|
||||
|
||||
A clean system uses:
|
||||
- working memory
|
||||
- session memory
|
||||
- durable memory
|
||||
- procedural memory
|
||||
- artifact memory
|
||||
|
||||
And it follows one hard rule:
|
||||
|
||||
Retrieval before generation.
|
||||
|
||||
If the agent can look something up in a verified artifact, it should do that before it improvises.
|
||||
|
||||
## The Five Layers
|
||||
|
||||
### 1. Working Memory
|
||||
|
||||
This is what the agent is actively holding right now.
|
||||
|
||||
Examples:
|
||||
- current user prompt
|
||||
- current file under edit
|
||||
- last tool output
|
||||
- last few conversation turns
|
||||
- current objective and acceptance criteria
|
||||
|
||||
Properties:
|
||||
- small
|
||||
- hot
|
||||
- disposable
|
||||
- aggressively pruned
|
||||
|
||||
Failure mode:
|
||||
If working memory gets too large, the agent starts treating noise as priority and loses the thread.
|
||||
|
||||
### 2. Session Memory
|
||||
|
||||
This is what happened during the current task or run.
|
||||
|
||||
Examples:
|
||||
- issue number
|
||||
- branch name
|
||||
- commands already tried
|
||||
- errors encountered
|
||||
- decisions made during the run
|
||||
- files already inspected
|
||||
|
||||
Properties:
|
||||
- persists across turns inside the task
|
||||
- should compact periodically
|
||||
- should die when the task dies unless something deserves promotion
|
||||
|
||||
Failure mode:
|
||||
If session memory is not compacted, every task drags a dead backpack of irrelevant state.
|
||||
|
||||
### 3. Durable Memory
|
||||
|
||||
This is what the system should remember across sessions.
|
||||
|
||||
Examples:
|
||||
- user preferences
|
||||
- stable machine facts
|
||||
- repo conventions
|
||||
- important credentials paths
|
||||
- identity/role relationships
|
||||
- recurring operator instructions
|
||||
|
||||
Properties:
|
||||
- sparse
|
||||
- curated
|
||||
- stable
|
||||
- high-value only
|
||||
|
||||
Failure mode:
|
||||
If you write too much into durable memory, retrieval quality collapses. The agent starts remembering trivia instead of truth.
|
||||
|
||||
### 4. Procedural Memory
|
||||
|
||||
This is “how to do things.”
|
||||
|
||||
Examples:
|
||||
- deployment playbooks
|
||||
- debugging workflows
|
||||
- recovery runbooks
|
||||
- test procedures
|
||||
- standard triage patterns
|
||||
|
||||
Properties:
|
||||
- reusable
|
||||
- highly structured
|
||||
- often better as markdown skills or scripts than embeddings
|
||||
|
||||
Failure mode:
|
||||
A weak system stores facts but forgets how to work. It knows things but cannot repeat success.
|
||||
|
||||
### 5. Artifact Memory
|
||||
|
||||
This is the memory outside the model.
|
||||
|
||||
Examples:
|
||||
- issues
|
||||
- pull requests
|
||||
- docs
|
||||
- logs
|
||||
- transcripts
|
||||
- databases
|
||||
- config files
|
||||
- code
|
||||
|
||||
This is the most important category because it is often the most truthful.
|
||||
|
||||
If your agent ignores artifact memory and tries to “remember” everything in model context, it will eventually hallucinate operational facts.
|
||||
|
||||
Repos are memory.
|
||||
Logs are memory.
|
||||
Gitea is memory.
|
||||
Files are memory.
|
||||
|
||||
## A Good Write Policy
|
||||
|
||||
Before writing memory, ask:
|
||||
- Will this matter later?
|
||||
- Is it stable?
|
||||
- Is it specific?
|
||||
- Can it be verified?
|
||||
- Does it belong in durable memory, or only in session scratchpad?
|
||||
|
||||
A good agent writes less than a naive one.
|
||||
The difference is quality, not quantity.
|
||||
|
||||
## A Good Retrieval Order
|
||||
|
||||
When a new task arrives:
|
||||
|
||||
1. check durable memory
|
||||
2. check task/session state
|
||||
3. retrieve relevant artifacts
|
||||
4. retrieve procedures/skills
|
||||
5. only then generate free-form reasoning
|
||||
|
||||
That order matters.
|
||||
|
||||
A lot of systems do it backwards:
|
||||
- think first
|
||||
- search later
|
||||
- rationalize the mismatch
|
||||
|
||||
That is how you get fluent nonsense.
|
||||
|
||||
## Recommended Data Shape
|
||||
|
||||
If you want a practical implementation, use this split:
|
||||
|
||||
### A. Exact State Store
|
||||
Use JSON or SQLite for:
|
||||
- current task state
|
||||
- issue/branch associations
|
||||
- event IDs
|
||||
- status flags
|
||||
- dedupe keys
|
||||
- replay protection
|
||||
|
||||
This is for things that must be exact.
|
||||
|
||||
### B. Human-Readable Knowledge Store
|
||||
Use markdown, docs, and issues for:
|
||||
- runbooks
|
||||
- KT docs
|
||||
- architecture decisions
|
||||
- user-facing reports
|
||||
- operating doctrine
|
||||
|
||||
This is for things humans and agents both need to read.
|
||||
|
||||
### C. Search Index
|
||||
Use full-text search for:
|
||||
- logs
|
||||
- transcripts
|
||||
- notes
|
||||
- issue bodies
|
||||
- docs
|
||||
|
||||
This is for fast retrieval of exact phrases and operational facts.
|
||||
|
||||
### D. Embedding Layer
|
||||
Use embeddings only as a helper for:
|
||||
- fuzzy recall
|
||||
- similarity search
|
||||
- thematic clustering
|
||||
- long-tail discovery
|
||||
|
||||
Do not let embeddings become your only memory system.
|
||||
|
||||
Semantic search is useful.
|
||||
It is not truth.
|
||||
|
||||
## The Common Failure Modes
|
||||
|
||||
### 1. One Giant Vector Bucket
|
||||
Everything gets embedded. Nothing gets filtered. Retrieval becomes mood-based instead of exact.
|
||||
|
||||
### 2. No Separation of Lifetimes
|
||||
Temporary scratchpad gets treated like durable truth.
|
||||
|
||||
### 3. No Promotion Rules
|
||||
Nothing decides what gets promoted from session memory into durable memory.
|
||||
|
||||
### 4. No Compaction
|
||||
The system keeps dragging old state forward forever.
|
||||
|
||||
### 5. No Artifact Priority
|
||||
The model trusts its own “memory” over the actual repo, issue tracker, logs, or config.
|
||||
|
||||
That last failure is the ugliest one.
|
||||
|
||||
## A Better Mental Model
|
||||
|
||||
Think of memory as a city, not a lake.
|
||||
|
||||
- Working memory is the desk.
|
||||
- Session memory is the room.
|
||||
- Durable memory is the house.
|
||||
- Procedural memory is the workshop.
|
||||
- Artifact memory is the town archive.
|
||||
|
||||
Do not pour the whole town archive onto the desk.
|
||||
Retrieve what matters.
|
||||
Work.
|
||||
Write back only what deserves to survive.
|
||||
|
||||
## Why This Matters for OpenClaw
|
||||
|
||||
OpenClaw-style systems get useful quickly because they are flexible, channel-native, and easy to wire into real workflows.
|
||||
|
||||
But the risk is that state, routing, identity, and memory start to blur together.
|
||||
That works at first. Then it becomes sludge.
|
||||
|
||||
The clean pattern is to separate:
|
||||
- identity
|
||||
- routing
|
||||
- live task state
|
||||
- durable memory
|
||||
- reusable procedure
|
||||
- artifact truth
|
||||
|
||||
This is also where Hermes quietly has the stronger pattern:
|
||||
not all memory is the same, and not all truth belongs inside the model.
|
||||
|
||||
That does not mean “copy Hermes.”
|
||||
It means steal the right lesson:
|
||||
separate memory by role and by lifetime.
|
||||
|
||||
## Minimum Viable Agentic Memory Stack
|
||||
|
||||
If you want the simplest version that is still respectable, build this:
|
||||
|
||||
1. small working context
|
||||
2. session-state SQLite file
|
||||
3. durable markdown notes + stable JSON facts
|
||||
4. issue/doc/log retrieval before generation
|
||||
5. skill/runbook store for recurring workflows
|
||||
6. compaction at the end of every serious task
|
||||
|
||||
That already gets you most of the way there.
|
||||
|
||||
## Final Recommendation
|
||||
|
||||
If you are unsure where to start, start here:
|
||||
|
||||
- Bucket 1: now
|
||||
- Bucket 2: this task
|
||||
- Bucket 3: durable facts
|
||||
- Bucket 4: procedures
|
||||
- Bucket 5: artifacts
|
||||
|
||||
Then add three rules:
|
||||
- retrieval before generation
|
||||
- promotion by filter, not by default
|
||||
- compaction every cycle
|
||||
|
||||
That structure is simple enough to build and strong enough to scale.
|
||||
|
||||
## Closing
|
||||
|
||||
The real goal of memory is not “remember more.”
|
||||
It is:
|
||||
- reduce rework
|
||||
- preserve truth
|
||||
- repeat successful behavior
|
||||
- stay honest under load
|
||||
|
||||
A good memory system does not make the agent feel smart.
|
||||
It makes the agent less likely to lie.
|
||||
|
||||
#GrepTard
|
||||
245
reports/greptard/2026-04-06-agentic-memory-for-openclaw.pdf
Normal file
245
reports/greptard/2026-04-06-agentic-memory-for-openclaw.pdf
Normal file
@@ -0,0 +1,245 @@
|
||||
%PDF-1.4
|
||||
%“Œ‹ž ReportLab Generated PDF document (opensource)
|
||||
1 0 obj
|
||||
<<
|
||||
/F1 2 0 R /F2 3 0 R
|
||||
>>
|
||||
endobj
|
||||
2 0 obj
|
||||
<<
|
||||
/BaseFont /Helvetica /Encoding /WinAnsiEncoding /Name /F1 /Subtype /Type1 /Type /Font
|
||||
>>
|
||||
endobj
|
||||
3 0 obj
|
||||
<<
|
||||
/BaseFont /Helvetica-Bold /Encoding /WinAnsiEncoding /Name /F2 /Subtype /Type1 /Type /Font
|
||||
>>
|
||||
endobj
|
||||
4 0 obj
|
||||
<<
|
||||
/Contents 17 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
5 0 obj
|
||||
<<
|
||||
/Contents 18 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
6 0 obj
|
||||
<<
|
||||
/Contents 19 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
7 0 obj
|
||||
<<
|
||||
/Contents 20 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
8 0 obj
|
||||
<<
|
||||
/Contents 21 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
9 0 obj
|
||||
<<
|
||||
/Contents 22 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
10 0 obj
|
||||
<<
|
||||
/Contents 23 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
11 0 obj
|
||||
<<
|
||||
/Contents 24 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
12 0 obj
|
||||
<<
|
||||
/Contents 25 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
13 0 obj
|
||||
<<
|
||||
/Contents 26 0 R /MediaBox [ 0 0 612 792 ] /Parent 16 0 R /Resources <<
|
||||
/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
|
||||
>> /Rotate 0 /Trans <<
|
||||
|
||||
>>
|
||||
/Type /Page
|
||||
>>
|
||||
endobj
|
||||
14 0 obj
|
||||
<<
|
||||
/PageMode /UseNone /Pages 16 0 R /Type /Catalog
|
||||
>>
|
||||
endobj
|
||||
15 0 obj
|
||||
<<
|
||||
/Author (\(anonymous\)) /CreationDate (D:20260406174739-04'00') /Creator (\(unspecified\)) /Keywords () /ModDate (D:20260406174739-04'00') /Producer (ReportLab PDF Library - \(opensource\))
|
||||
/Subject (\(unspecified\)) /Title (\(anonymous\)) /Trapped /False
|
||||
>>
|
||||
endobj
|
||||
16 0 obj
|
||||
<<
|
||||
/Count 10 /Kids [ 4 0 R 5 0 R 6 0 R 7 0 R 8 0 R 9 0 R 10 0 R 11 0 R 12 0 R 13 0 R ] /Type /Pages
|
||||
>>
|
||||
endobj
|
||||
17 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 1202
|
||||
>>
|
||||
stream
|
||||
Gatm:;/b2I&:Vs/3++DmOK]oT;0#utG&<!9%>Ltdp<ja\U+NO4k`V/Ns8*f_fh#+F:#]?,/b98)GB_qg7g]hl[/V`tJ2[YEH;B)I-rrSHP!JOLhA$i60Bfp08!),cE86JS,i\U@J4,<KdMrNe)8:Z-?f4fRGWMGFi5lCH&"_!I3;<":>..b&E%;l;pTQCgrn>He%kVtoVGk'hf)5rpIo-%Y5l?Vf2aI-^-$^,b4>p^T_q0gF?[`D3-KtL:K8m`p#^67P)_4O6,r@T"]@(n(p=Pf7VQY\DMqC,TS6U1T\e^E[2PMD%E&f.?4p<l:7K[7eNL>b&&,t]OMpq+s35AnnfS>+q@XS?nr+Y8i&@S%H_L@Zkf3P4`>KRnXBlL`4d_W!]3\I%&a64n%Gq1uY@1hPO;0!]I3N)*c+u,Rc7`tO?5W"_QPV8-M4N`a_,lGp_.5]-7\;.tK3:H9Gfm0ujn>U1A@J@*Q;FXPJKnnfci=q\oG*:)l<j4M'#c)EH$.Z1EZljtkn>-3F^4`o?N5`3XJZDMC/,8Uaq?-,7`uW8:P$`r,ek>17D%%K4\fJ(`*@lO%CZTGG6cF@Ikoe*gp#iCLXb#'u+\"/fKXDF0i@*BU2To6;-5e,W<$t7>4;pt(U*i6Tg1YWITNZ8!M`keUG08E5WRXVp:^=+'d]5jKWs=PEX1SSil*)-WF`4S6>:$c2TJj[=Nkhc:rg<]4TA)F\)[B#=RKe\I]4rq85Cm)je8Z"Y\jP@GcdK1,hK^Y*dK*TeKeMbOW";a;`DU4G_j3:?3;V%"?!hqm)f1n=PdhBlha\RT^[0)rda':(=qEU2K/c(4?IHP\/Wo!Zpn:($F"uh1dFXV@iRipG%Z''61X.]-);?ZT8GfKh,X'Hg`o<4sTAg2lH^:^"h4NUTU?B'JYQsFj@rEo&SEEUKY6(,aies][SXL*@3>c)<:K0)-KpC'>ls)3/)J=1GoN@nDCc'hpHslSSGWqRGNh']0nVVs9;mm=hO"?cc/mV08Q=_ca/P`9!=GEeSU3a4%fS?\Li@I93FW5-J+CH_%=t*SpX*>A"3R4_K$s0bi&i?Nk\[>EcS,$H:6J,9/Vb&]`cFjMu(u>)9Bg;:n?ks43,alko`(%YTBIJF]a0'R^6P\ZBaehJA&*qOpGC^P5]J.,RPj?'Q\.UFN>H.?nS)LMZe[kH6g38T.#T*LC_lG'C~>endstream
|
||||
endobj
|
||||
18 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 926
|
||||
>>
|
||||
stream
|
||||
Gau1-9lJc?%#46H'g-ZWi*$')NWWHm^]plBNk6\t)lp@m=-VJ$hiDeA<cinMegV47NbRf&WL";&`;C4rDqtMC?,>\S!@!*F%>ZRX@.bNm=,Zs0fKWNW:aXAab4teOoeSs_0\e>@l*fD!GY)nNUeqW&I`I9C[AS8h`T82p%is)b&$WX!eONEKIL[d+nd%4_mV]/)Wup,IMr16[TcU=)m9h3H0Ncd70$#R6oC-WsLG8"JWS".'1J?mc4%TpP0ccY/%:^6@Lblhan.BE1+4-0mb%PaheJ.$*bN4!^CY#ss48"+HFT\qPEH"h-#dmYBXcbt'WZm>$'11!&KAlJirb9W-eu9I]S7gLenYQ^k0=ri-8<S7Oec`CEa76h8)b#B&\aD/)ai\?>7W(+"M-)"YQ>:s!fE?Ig(8]+Z;.S@rn9Rr:8_&e9Tf3DbAWcM[]bU,*"s/c;;gJO/p;UuYK8t=0i%h\Zquj1a3na;>+XaD$=lbJ%(UR&X2W=ig_kj]1lDZRm1qi!SI^4f^Y/aP,(FKi^<nZ>K^PG9%nmVof.,pCO5ihrG`W&g'@SB%_;hW+(@1pC0^QmS`IS:?.r(5'k3`XsL^;'^E%'Ni'^u*153b[V/+H$PpdJ=RR1`b;5PB7&L!imo?ZSX8/ps`00MM'lYNm_I+*s$:0.n)9=kcnKi%>)`E*b]E$Tsp\++7'Y40'7.ge+YD>!nhk$Dn.i,\5ae:#gG]1[DiiPY0Ep@\9\/lQh,/*f#ed>5qa1)Wa"%gLb,Qo@e''9^VhTr"/"<+BLOAEjAc)8r*XcY_ccKK-?IHPL6*TsYd1]lBK$Lu\5e0nI``*DkQ1/F/.\[:A(Ps&&$gB8+_;Qlo?7b^R_7&2^buP18YSDglL,9[^aoQh1-N5"CTg#F`#k)<=krf*1#s<),2]$:YkSTmXTOj~>endstream
|
||||
endobj
|
||||
19 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 942
|
||||
>>
|
||||
stream
|
||||
Gau0BbAQ&g&A7ljp6ZVp,#U)+a.`l:TESLaa'<:lD5jC#Kr")3n%4f9UMU.2Yn4u1CbE-%YMkR/8.QZRM[&*<%X/El(l*J@/.Q.1^VYE5GZq?Cc8:35ZQQ+3Zl0FTHFogdfu7#,,`jr4:SI[QHoXt]&#,B'7VGbX+5]B`'CtnCrssGT_FRb/CGEP>^n<EiC8&I5kp+[^>%OC(9d^:+jXoC3C#'-&K2RW0!<FL('%Wf0d@MAW%FeRV`4]h9(b7-AhKYAYY`)l'ru+dY2EWPm``\J-+CJoNd:2&50%9!oSMK6@U*6^^a=@VUF0EPd,((@#o#;=rK5`gc%j)7aYA@NB-0]KLKEYR'R%pq=J>lL$9@-6&?D@^%BP#E?"lh6U9j,C^!_l^jiUqcYrt8$Rd<J/4anQ$Ib4!I(TAIUBi9$/5)>&Z(m5/]W@p>VrJgKA<0H*7/q*l&uh'-ZKOSs^Zk?3<R4%5BJpXi[$75%c1c3(,::20$m<bO$)U6#R?@4O!K]SpL_%TrFLV\Kr5pb%;+X1Io_VDC_4A'o't[p)ZLC13B^\i!O_`J_-aM:kH]6("%#L=[+$]682Hq?>$[eE7G'\gd'#2X#dLW26gCW3CAGQX1)8hn1cM13t,'E#qDIDlXCq+aX@B9(,n)nMHUolD*j]re<JYZd=cL17qAb<=]=?>6Lu@1jr45&$1BR/9E6?^EpTr?'?$sGj9u._U?OOV<CHZ!m!ri`"l-0Xf],>OlI7k\$*c<_Mr&n'7/)N@[jL4g;K1+#cC(]8($!D=4H71LjK<:K]R^I3bPLD:GnWD7`4o1rlB@aW<9X7-k^d)T*]0cp-lp`k*&IF3(lcZ)[SK^UC4k;*%S:XlI`Vgl(g;AQ.gME?L%/f^idEJ]!4@G^"Z)#nD[;<B>K_QW8XJOqtA"iZ>:SL771WKcgnEc&1i84L#qlG~>endstream
|
||||
endobj
|
||||
20 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 987
|
||||
>>
|
||||
stream
|
||||
Gatm:;/ao;&:X)O\As09`(o<&]FB]@U`fbr`62KB`(<[e\?d3erd2q)1Vj?/CJ^fVoj.LOh.M4]'I%qgpfhnAmg=;\9n>&JC7kl)TX]RI`Th>0:S'\#N)@I>*EUX\5&:gp'T*Abo,itH7"1sR*?m0H>hmaY"7&?^#ucC28i.0(_Du=VqZ;ZD:qZFF?h!k31Wi7[GgJqbSkk*QeV#^teuo)p6bN21oM-=&UjX3!jue'B3%JD^#:nB-%"]bp16@K12XLO'CPL7H7TMf3Ui6p7Y+=of/Nn.t/9JaF/%tDfDH5Fpcj"<#eBV39)ps<)f!;;ThZR*E;,3j5L?17a%*rS^>,W5&!M-B(OQVl"ZU(a%0?"^n_)g6m$/.oX[4!d0p#)LoVNLYfd<fgNp=a<hHuuU[p"ick(M7b?7Ghm9-Y=`["$;aD[$Cii:)SoA>g6B"1>Ju;AMiM85U_[K,bFeG3WCnO@sSPs4=8+hjAH%\GYNQHn4@fW*.e3bDPVY,T]C,K4MSVL7TiR%<(Q'e!pII'<QX86En^fAPiNFE4';kSXZo%Ip\1E:[Jpf!,gN=dcamf4g-Gor9g\Y"K\b"`Gi8!$`W^p&jDP?$V9AB-)-aItX2F38VpV7;SItfle:KAj)<7!$@P)D`oJg#DHE$dF2,L>3N5P3tS<nITKDT;G7!!dIV3>>]=D7"cFZXGZlL=Z8AE23M/P@g#$-IP>@lo&,`uaM(oak.<(2&<F8ICC8PMpGRe*M"X^Q(k'Eti78p2KQ,L4^PO_);p9=%tof'esFm8I0=)ntQn&YdN7A()ts&IV\F9!Vo*O_q8B_ogb_JloTL]?MWs^fWVtemfq1J&'>rQ!Gl^h-rl=."$\:BVfXTG@qQ0MLZXpKSSLl_:PS$Gqc3'kc[Y3\i<YV5CnM`3Osf:ooPC$.b":P&i)=Ua=Ik@kmI0jL'11ie\c.RuE5qpu4E1NH['&>V_<g-eDH6LWTRbQgGN/NO~>endstream
|
||||
endobj
|
||||
21 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 976
|
||||
>>
|
||||
stream
|
||||
Gatn%9lHLd&;KZQ'm"2lFMmZ\kbW$_[#.h^8qN:#0,kbPf"g"OlaZAdmf9d2\S/&%aqb0cn[tH]I:kQbo\oa'DZHpq\GX3pqiI)Y6P`#^%;rK)HH,\T`ZEA&PU.J95J&u`G_a(\k4Q4V@RdQ^7nUQ@aI7\=FlsdAA%]@h;JCfdQ<(F%BWt?[G6,Q35J2^:-Y2[,*I"F&311kNA#/)N06me2nE'tJcf%aP2:tM>BS<dlTb_bJk[_]\H-BIpdXna`WCAfq%/pWKKt/aGmUl:m4P"/mG-E?IB@MUP_T@_aoK;!<68JUW73*UW-oSY0l*5Mu#1_;:/nE<GWTZ:m_WKB@r'O^%1G9V[nL-R(?Jjb7@%gO#@Y(ZK)kHWQUl3rf;CA+Pnek"R18hK5S?*j6&.2R+W3OSZW9MnQ;+jQmC6e0=>"_q7Ic.KH+%qS0mfVknj$&O`'GunE3E;JiQV%+ae3U#D4Qp@rqa>l"&p97997.L4I+JO:Q`)V2=VQpQ$Km2[la-7d@:'f*JgDK?Tf!S+3;k7a&iS<"@BdNHH5W5=?=CQ1BlBmV*`&X.#?pkg09=;rOt4,"5oNKE"q:-#Br$r]$;Sc3BIc`<>N:B7E@4)j(XSFJ3DsnF>acsu"#i%,VD9ASfTGRtMG+#lM@`C>pmu))6\9tg/PSGW5F=6FD"n54&=DGb_NJ-O,25mZj0X?P$^a00jaM4U9QA+A/4c%6G/e!$TMW>6MgW&M\o9;a5NYK*UgZSOJ9D6qeAaO06$aTmT[7sACbhM'WodG,l7H2LAF@4;CH-"'BtDFLKMl*N0l,so+Y^11B[Tjp$Tkbi`j\dqRr/G=W\m=SB=%+fAb.Wlk@(_.S3(ZW0iq)%D1Mjq$S1//&hBm9n^.Zaq8=9/Q@3MV^%7@.On$P`k+6Bi23KZJ(\7\d#)Bml=jb`BY)"oCrobCdgtt>C82IdO77,t,RgjJ8mJD__R/I%aB^5$~>endstream
|
||||
endobj
|
||||
22 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 892
|
||||
>>
|
||||
stream
|
||||
Gatn%9lnc;&;KZL'mhK*X'5(,30ol(Zia69DHmm&*Lf#d.`lPV?]PmK)(6A6LbkXELEu1gbO<'T'EWX/ok1N0\B<e$'*ZN$YCK(fK)?[-o$QKRDH8(bUl:JT0"7UlH;r-Yp-uXJ.jQk1MghS>rH<OuS^)[tKJ/O-?U60/4PVk"Tp"]Y432n;rn`bYm.D/H)3;86I?8p<5>g1&i9Vgc;^a]':`(lkcTrLfcq@$\6*s,I%PO`;MkUEY[4E56)C`$0)TRK'puELcp=^#`_a+iXFIe/djYK,VdQsB9bAN.ja-@GSB>\8RE7sds/<r%o\2WK&V(p(97sd<0D^YPA$[LQAWu#CK0QeUPg4"#DMbY4Y,2`h^%.T&e0bS$o_P5^qic.@7UN&$n6B`P"YnHdi9E#C-&!I#'f0LbL+Kl&umM9&4Hq.N6Bo?pXjE$7@$t3V@nsV!G-bD'maV5,ck,!G@k>9;fT*#m@D_QSnCIDmD6Q`4e:/%LSHSlYZC`)4c?U'sF&I-,i>SVDA1u9[gjsh9t-lk`B2@S8Q<#69&XJVQ7UbZ7_QmKXpEf%qN=\H*!BiH=iWXMfq6FOol@D&-jM4&/B)nd"=T@j@L@4Ft\!jMkQmD8;?lg?IN8=]%)dh_(*3JG(0&t#=*#i(:M?[U8*1##!TnT*0fm=i@m"1fj$E\L.=*UkIW[*i<[=Hj6s(gH*ETphfbhM`bu35Ut059Yi;&_9P(b-Pp^+I]QDTL7Cm-5kK<ctKd(+6Le)gX<+KV?hS//]aqFZEUDFf@YFmP>%dV$Z$;/g1sS)`['3g),T&l"jnbmH;3=00u^G?$1=Cg;#(0uD,G7_6fMp>>ET1>g6HM`JO>F!4d<jTHFcHc-'#!PI9kOX;t.5D_h$3RR5jTI~>endstream
|
||||
endobj
|
||||
23 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 1105
|
||||
>>
|
||||
stream
|
||||
Gatm:bECU<&A7<Zk3/Vl#XRr'*PJ_K8sPs&QPEkT_3..m&P6X:cl;q3#(3B92%.Mu/sT`[9&8juBn/Q=%mI^n*PcLm5J?0o@jl*M#tpq9#B,LE_hQK=AeF)YB,OuMX)pKhs+Oi0'HR.rs(M>"aKO+R"VI++?@:aX%T[_PPt&:t=Ho]<rrC$q;#KO.rhRS9@+fFA'q!`c-_2`\C^9:QW4iSTjnVnqXmV[F:9#'ZAP;t>bp`pb$+<PjhU=fs'HT.?`%%N8\P2?r][kGF#;&!PmN/Edmb,H?QF]FCl+qo<gCn\]BgXI3OXqL\Z4mn/[VIGi;KtCK46fK-)`)a>P_Hq_4g[L)l#qD.iigX[G\NZcgcg[\-VVH/\"'>C>^4mT*qM5C!TR=I;n&.o]$7sc1_GUrMigamdOFkq5O5K$4hN)Y$jP\(Y0[AeM2\:a)D*#VKkkCB#/Bm<^F8@Wr.Uur#&]4eJ1a5@fgTQOkP""sT+\U\QU6>McDJR<_/K1.j]]&K^OGt\3hu2NChH[Nkd7L!hFibAW1No</'p035I,CI2CEdier6/q1#%-f2.M+LE/-qt.#"VM7-gDDdA\bRl%E[:6D#(H*OV#Y[S&q=pICBVPgC;4N\kM$!MLEj]Pm=i$q%mQ(9OEtfSR.XX\N?TkmsON4j*D*BZ'\&S=cLI/'qb+$;3ei#EO5r-@q1%!E,kn&!Tc=H89P>Wo!!HC=??_2Nd,Q??;GE@P1n9>;F>j.<6]3'@e3GcHiH8[M3<'Zr+U>nS"UOZ>$t+\uib/EY[*X4A&QJGGL'*8e^Z6QEJ2BS;XpsXYf8jbq%gR:"k]:PkIV-+KLa!_(SZaU\Ja*4B\tQU8NJ,iDU_SaXm'5!IlBaLCt_-"!s>NUV<FMaUWb4-0;5=Ti4?hDh*GKe:"HfY`p>Eq:_,Go;-EJh9\QsdQ4>iXbh3rc['8.Ks[q.%'s-^$bhV77r)JJ/NVSni'$do2"]O7)e-^kN5_iNP,3,S7]J]J4J1Y">*)RD`GW[OL*Z^-@?J'U=gAeSS1fR(O.dZ'3V_iDP&"eRA_eM#Lc4e(@.0ZijofJ,rf?[4p,^jX?Y/d0]@V.rI#8<$IfZ<4,)Q~>endstream
|
||||
endobj
|
||||
24 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 1129
|
||||
>>
|
||||
stream
|
||||
GatU2bHBSX&Dd46ApIV9W1pA[]1_uc$hU/f.\rPMBR+-nVF4^QZE(b/:lcg2<>\,Y%E%>T3eoM(cB(=_0/e9F%D\G7^AF<!MkI#!`BapODt$]1Heu#QB,X)PYoon1ZqC5KIui4e#o!jIc2D@(9^#NWEC1$.$kFF^QGFYHQ$DROQdB-8oDlaYXV#GC?VIT5i#=*DL>lEmr0CZ]TVmR_?0JK"'brCh$Z)k]/T>"Hiei3_4:1T2&U50aQb`31_-Ei30tE//_iG..AYE&9Sha.nq'd[fX]8ltK]_9,)"0BsH&#E.]K-7e;T,\+D>\(CL95-=;8KpV:T2p8+0L;d3:cW,\WapQ,"`pA.oOV,QsO.7<:(r,K.pZ3G*5=9i-?-CLaD9d!g\QYd1+mW4T.LrM.m*/5OqJqT(N(P9eq*bZ43)In9]rX&!Gh_gu:HK7r-nYF/Qh:ZGs2rVJSVJAaDZ#1kW'c(c\:EhI+l,Gj\"GTnFJljL!u93KQoH.Z+1]UVoYNCYlKJ?a\ZeL*(uU-U;PRQhHoq\/ag#3)s`>.r`a?8TjX*/I@8N\oQpm?NOT<PW-8r4%fs&RJ_T"@P#",>cZ_=pA>3UKWPu4QjtpA#Aqo?*U5%Yk:4VPNS8`236=)m/KD%C-%Wd065pl*G-D-Y>rbkOau6OiSc,RKj#C-SFWAZl>Gr^0&pXl@.-#JE,W-H_>Z2uap[SWc"a?0.0=C=Ylq&>o@*Ct,6;VCbJWS1?/LM-jiq#M(e%;:.pn^`VFmMP+nU5]#hMb:e4SHJOM@TA0JM5L.lJC)uV!JYGCNDU1QGAe=+P"r191)0=<e0ZIC=e_]RV9f"CHgQ5N7FpkUL)ZngE4g,gJ#`F/%BtPeM=e*D0^u!#pU+G7#T@;)JZ9lCSOWQ]lk#Wq]K)6P:0Z;)nk[8.!:PYj.,35!L9Z(k*uYJ48/2R3SOuBl\%iQC>F!#`l#?q+OM%f0,Gi:D8ffYEiIZ1+QoHdEM]Du;H.<uj0s4J-UG%]0q`m9*4Yakng%DdoF8GL@rK+G3s.+oBd3MRV-B:Q@@.5Og)//&oaMY:?r0#AGWSW@,FXXMA`67J.\YR1X:&*,6$h-r3\2j)mjlFba:iD%``q)3p']OO$^k'_f)~>endstream
|
||||
endobj
|
||||
25 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 1016
|
||||
>>
|
||||
stream
|
||||
Gatm;9lo#B&A@Zcp6`:f@`_um0qDo'SP2+Z3T_OP27U^Cd75Vb^+0J'Rki.)A3>LLf'#9a54'"''"[l+Zg%NCB5hk0JI@i&^b_:mlioZ"7e\/,ph#XR.6&jAFW*ttULl6T+7c>?;O55%gWu+;mpW@Qdhue+4/ip1Zrsg8!\3"fr8leOlmL$6#H1k<rdcRHPG)rIi6^1YpE-N$%cPIq9hVsm<>^E@bFt"HKA%8)<sY?RX'0Ffkd<bc++Q>(nLm,@I>(-kc:0hTq+qP=K@X4e(U^CuP[8l+B3K\E$KfuSK<W-4NHCDP."sJu8g1+VY?IbTN:5q5YA!;(Qb\GOjc'Sc%jG\OhSf$$#HOt7Up"Z1RAP[iFPrGjf[lc8^]u]#;MSto\)=&f`CT'RP)Zf`id%1Z"C=lB[NnIC)3jf[.cN!q[>.L](C15J)n?E%K.KWtaB12]7P7=T"8=%["j`)50O?"N0kTBa>l7BA:K@HkG>sJ@eZ6,+nU*i!B1E8U;)u08==T>.8e-F(kNf_4,tO2ZHuD1(2Bb$;FP'a9SaUJ.#,a2!fgN'W35R9u%tXVQV"R4UVKQ&DSDE_@KM,[SciKceT&2pbjN3l6M(&b,I9F@R(r$A`3dka]06XRYCAep8fbE",=%L+D"\ctiRfMSL&t*NB>[U_^m+B$Fo>gAE4TVN\eMU@W+G0+jD,e\-'m=uAOp>/X9!pecQ3u@1?!En=K,m$1kJ8O`@uZK?.XEEQ<9[?>s0@?l&QIL7IO#INB?;k5&G[J'ciL4(^2fp<d6>!U[oU>@ZsP@OB:Jd!eKDu@kWMY;q'T]'WT)2GdZTGs$5G[O%%QSkT9QeFjY7O@%WM4u1Z7@<0PI5CG"K+M9crG_*HmkY8N@27\e?8F87Q]tA?"X_1m1:1k2fu+f8agFgQ\W3e*I22C$ht*jD\di,#N&M6<[<S7IrdYC:mcjTI0Td26ATL!+/`J8oK+@5th%#C(pV3E#L2H'"5$o4jl@6pGM2&Bj!YfK[F`%h]J3~>endstream
|
||||
endobj
|
||||
26 0 obj
|
||||
<<
|
||||
/Filter [ /ASCII85Decode /FlateDecode ] /Length 257
|
||||
>>
|
||||
stream
|
||||
Gat=dbmM<A&;9LtM?,Bq+XUCIH;m]Q]Eie6@^l$.0q.1O\$nRO;FF=sQ=X^]Dcd<CQ#-;'!t0h18j\Q31<=tN:g5Ic].s,#L'#bI`oE`_ZX$'4:Z9;]ZaMlp:<FAa_P^5AAVbZ8.5dkb7rD!-:"n54Y&p6l;F:p1j.VYm_&iqS:)AtMW.<?Rom?a^Jf<2`GMPAi*;uF@dmDk0[34X/*7%TKn>;Z<4$<q\Ld/O$S3DYaA+eE=Xt$\jLCA>2IHYJN~>endstream
|
||||
endobj
|
||||
xref
|
||||
0 27
|
||||
0000000000 65535 f
|
||||
0000000061 00000 n
|
||||
0000000102 00000 n
|
||||
0000000209 00000 n
|
||||
0000000321 00000 n
|
||||
0000000516 00000 n
|
||||
0000000711 00000 n
|
||||
0000000906 00000 n
|
||||
0000001101 00000 n
|
||||
0000001296 00000 n
|
||||
0000001491 00000 n
|
||||
0000001687 00000 n
|
||||
0000001883 00000 n
|
||||
0000002079 00000 n
|
||||
0000002275 00000 n
|
||||
0000002345 00000 n
|
||||
0000002626 00000 n
|
||||
0000002745 00000 n
|
||||
0000004039 00000 n
|
||||
0000005056 00000 n
|
||||
0000006089 00000 n
|
||||
0000007167 00000 n
|
||||
0000008234 00000 n
|
||||
0000009217 00000 n
|
||||
0000010414 00000 n
|
||||
0000011635 00000 n
|
||||
0000012743 00000 n
|
||||
trailer
|
||||
<<
|
||||
/ID
|
||||
[<25b005833ac6719201eda8c8a8690d7b><25b005833ac6719201eda8c8a8690d7b>]
|
||||
% ReportLab generated PDF document -- digest (opensource)
|
||||
|
||||
/Info 15 0 R
|
||||
/Root 14 0 R
|
||||
/Size 27
|
||||
>>
|
||||
startxref
|
||||
13091
|
||||
%%EOF
|
||||
@@ -39,8 +39,14 @@ BACKEND_GROQ = "groq"
|
||||
BACKEND_GROK = "grok"
|
||||
BACKEND_KIMI = "kimi-coding"
|
||||
BACKEND_OPENROUTER = "openrouter"
|
||||
BACKEND_OLLAMA_HERMES3_8B = "ollama-hermes3-8b"
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST = "ollama-llama3-1-latest"
|
||||
BACKEND_OLLAMA_QWEN2_5_14B = "ollama-qwen2-5-14b"
|
||||
|
||||
ALL_BACKENDS = [
|
||||
BACKEND_OLLAMA_HERMES3_8B, # Prioritize local Ollama models
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST,
|
||||
BACKEND_OLLAMA_QWEN2_5_14B,
|
||||
BACKEND_ANTHROPIC,
|
||||
BACKEND_OPENAI_CODEX,
|
||||
BACKEND_GEMINI,
|
||||
@@ -243,22 +249,28 @@ class TaskClassifier:
|
||||
# Order matters: first is most preferred
|
||||
TASK_BACKEND_MAP: Dict[TaskType, List[str]] = {
|
||||
TaskType.CODE: [
|
||||
BACKEND_OPENAI_CODEX, # Best for code generation
|
||||
BACKEND_ANTHROPIC, # Excellent for code review, complex analysis
|
||||
BACKEND_KIMI, # Long context for large codebases
|
||||
BACKEND_GEMINI, # Good multimodal code understanding
|
||||
BACKEND_GROQ, # Fast for simple code tasks
|
||||
BACKEND_OPENROUTER, # Overflow option
|
||||
BACKEND_GROK, # General knowledge backup
|
||||
BACKEND_OLLAMA_HERMES3_8B, # Local, good for many code tasks
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST,
|
||||
BACKEND_OPENAI_CODEX, # Best for code generation
|
||||
BACKEND_ANTHROPIC, # Excellent for code review, complex analysis
|
||||
BACKEND_KIMI, # Long context for large codebases
|
||||
BACKEND_GEMINI, # Good multimodal code understanding
|
||||
BACKEND_GROQ, # Fast for simple code tasks
|
||||
BACKEND_OLLAMA_QWEN2_5_14B,
|
||||
BACKEND_OPENROUTER, # Overflow option
|
||||
BACKEND_GROK, # General knowledge backup
|
||||
],
|
||||
TaskType.REASONING: [
|
||||
BACKEND_ANTHROPIC, # Deep reasoning champion
|
||||
BACKEND_GEMINI, # Strong analytical capabilities
|
||||
BACKEND_KIMI, # Long context for complex reasoning chains
|
||||
BACKEND_GROK, # Broad knowledge for reasoning
|
||||
BACKEND_OPENAI_CODEX, # Structured reasoning
|
||||
BACKEND_OPENROUTER, # Overflow
|
||||
BACKEND_GROQ, # Fast fallback
|
||||
BACKEND_OLLAMA_HERMES3_8B, # Local reasoning
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST,
|
||||
BACKEND_ANTHROPIC, # Deep reasoning champion
|
||||
BACKEND_GEMINI, # Strong analytical capabilities
|
||||
BACKEND_KIMI, # Long context for complex reasoning chains
|
||||
BACKEND_GROK, # Broad knowledge for reasoning
|
||||
BACKEND_OPENAI_CODEX, # Structured reasoning
|
||||
BACKEND_OLLAMA_QWEN2_5_14B,
|
||||
BACKEND_OPENROUTER, # Overflow
|
||||
BACKEND_GROQ, # Fast fallback
|
||||
],
|
||||
TaskType.RESEARCH: [
|
||||
BACKEND_GEMINI, # Research and multimodal leader
|
||||
@@ -268,6 +280,9 @@ class TaskClassifier:
|
||||
BACKEND_OPENROUTER, # Broadest model access
|
||||
BACKEND_OPENAI_CODEX, # Structured research
|
||||
BACKEND_GROQ, # Fast triage
|
||||
BACKEND_OLLAMA_HERMES3_8B, # Local for basic research
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST,
|
||||
BACKEND_OLLAMA_QWEN2_5_14B,
|
||||
],
|
||||
TaskType.CREATIVE: [
|
||||
BACKEND_GROK, # Creative writing and drafting
|
||||
@@ -277,15 +292,21 @@ class TaskClassifier:
|
||||
BACKEND_KIMI, # Long-form creative
|
||||
BACKEND_OPENROUTER, # Variety of creative models
|
||||
BACKEND_GROQ, # Fast creative ops
|
||||
BACKEND_OLLAMA_HERMES3_8B, # Local for creative drafting
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST,
|
||||
BACKEND_OLLAMA_QWEN2_5_14B,
|
||||
],
|
||||
TaskType.FAST_OPS: [
|
||||
BACKEND_GROQ, # 284ms response time champion
|
||||
BACKEND_OPENROUTER, # Fast mini models
|
||||
BACKEND_GEMINI, # Flash models
|
||||
BACKEND_GROK, # Fast for simple queries
|
||||
BACKEND_ANTHROPIC, # If precision needed
|
||||
BACKEND_OPENAI_CODEX, # Structured ops
|
||||
BACKEND_KIMI, # Overflow
|
||||
BACKEND_OLLAMA_HERMES3_8B, # Prioritize local fast ops
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST,
|
||||
BACKEND_OLLAMA_QWEN2_5_14B,
|
||||
BACKEND_GROQ, # 284ms response time champion
|
||||
BACKEND_OPENROUTER, # Fast mini models
|
||||
BACKEND_GEMINI, # Flash models
|
||||
BACKEND_GROK, # Fast for simple queries
|
||||
BACKEND_ANTHROPIC, # If precision needed
|
||||
BACKEND_OPENAI_CODEX, # Structured ops
|
||||
BACKEND_KIMI, # Overflow
|
||||
],
|
||||
TaskType.TOOL_USE: [
|
||||
BACKEND_ANTHROPIC, # Excellent tool use capabilities
|
||||
@@ -295,15 +316,21 @@ class TaskClassifier:
|
||||
BACKEND_KIMI, # Long context tool sessions
|
||||
BACKEND_OPENROUTER, # Overflow
|
||||
BACKEND_GROK, # General tool use
|
||||
BACKEND_OLLAMA_HERMES3_8B, # Local tool use
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST,
|
||||
BACKEND_OLLAMA_QWEN2_5_14B,
|
||||
],
|
||||
TaskType.UNKNOWN: [
|
||||
BACKEND_ANTHROPIC, # Default to strongest general model
|
||||
BACKEND_GEMINI, # Good all-rounder
|
||||
BACKEND_OPENAI_CODEX, # Structured approach
|
||||
BACKEND_KIMI, # Long context safety
|
||||
BACKEND_GROK, # Broad knowledge
|
||||
BACKEND_GROQ, # Fast fallback
|
||||
BACKEND_OPENROUTER, # Ultimate overflow
|
||||
BACKEND_OLLAMA_HERMES3_8B, # Default to local first
|
||||
BACKEND_OLLAMA_LLAMA3_1_LATEST,
|
||||
BACKEND_ANTHROPIC, # Default to strongest general model
|
||||
BACKEND_GEMINI, # Good all-rounder
|
||||
BACKEND_OPENAI_CODEX, # Structured approach
|
||||
BACKEND_KIMI, # Long context safety
|
||||
BACKEND_GROK, # Broad knowledge
|
||||
BACKEND_GROQ, # Fast fallback
|
||||
BACKEND_OPENROUTER, # Ultimate overflow
|
||||
BACKEND_OLLAMA_QWEN2_5_14B,
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user