This release includes 2 breaking changes for platform teams planning a safe upgrade.
✓ No known CVEs patched in this version
Summary
AI summaryEmbedding model defaults changed from MiniLM-L6-v2 to MiniLM-L12-v2 and reranker enabled with an updated model.
Full changelog
What's new in v3.7.1
New defaults
| Setting | Old | New |
|---|---|---|
| Embedding model | all-MiniLM-L6-v2 | all-MiniLM-L12-v2 |
| Reranker | disabled | enabled |
| Reranker model | ms-marco-MiniLM-L-6-v2 | ms-marco-MiniLM-L-12-v2 |
MiniLM-L12-v2 — same speed class as L6, better quality, ~130MB.
MS MARCO MiniLM-L-12-v2 reranker — cross-encoder trained on passage retrieval. Re-scores the top candidates after initial vector+BM25 retrieval for significantly better search precision. Adds ~50ms latency per search.
Both models are downloaded to the persistent /data/models volume on first container start (not baked into the image).
Breaking Changes
- Embedding model default changed from `all-MiniLM-L6-v2` to `all-MiniLM-L12-v2`
- Reranker feature is now enabled by default and uses model `ms-marco-MiniLM-L-12-v2` instead of being disabled
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About dl4rce/flaiwheel
Self-hosted memory and governance layer for AI coding agents. 28 MCP tools with structured knowledge capture, hybrid search (semantic + BM25 + cross-encoder reranking), behavioral documentation nudges, cold-start codebase analyzer, and git-native storage. Single Docker container, zero cloud dependencies.
Related context
Beta — feedback welcome: [email protected]