d61dcfb83e
- openmemory: use qwen2.5:1.5b instead of gemma3:1b for fact extraction - test_pipeline.py: check qwen2.5:1.5b, fix SSE checks, fix Qdrant payload parsing, relax SearXNG threshold to 5s, improve marker word test - potential-directions.md: ranked CPU extraction model candidates - Root cause: mem0migrations collection had stale 1536-dim vectors causing silent dedup failures; recreate both collections at 768 dims All 18 pipeline tests now pass. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
14 lines
843 B
Markdown
14 lines
843 B
Markdown
# Potential Directions
|
||
|
||
## CPU Extraction Model Candidates (mem0 / openmemory)
|
||
|
||
Replacing `gemma3:1b` — documented JSON/structured output failures make it unreliable for mem0's extraction pipeline.
|
||
|
||
| Rank | Model | Size | CPU speed | JSON reliability | Notes |
|
||
|------|-------|------|-----------|-----------------|-------|
|
||
| 1 | `qwen2.5:1.5b` | ~934 MB | 25–40 tok/s | Excellent | Best fit: fast + structured output, 18T token training |
|
||
| 2 | `qwen2.5:3b` | ~1.9 GB | 15–25 tok/s | Excellent | Quality upgrade, same family |
|
||
| 3 | `llama3.2:3b` | ~2 GB | 15–25 tok/s | Good | Highest IFEval score (77.4) in class |
|
||
| 4 | `smollm2:1.7b` | ~1.1 GB | 25–35 tok/s | Moderate | Use temp=0; NuExtract-1.5-smol is fine-tuned variant |
|
||
| 5 | `phi4-mini` | ~2.5 GB | 10–17 tok/s | Good | Function calling support, borderline CPU speed |
|