mirror of
https://github.com/Crosstalk-Solutions/project-nomad.git
synced 2026-10-17 13:19:18 +02:00
94fb876f1b
Adds a `rag.embeddingModel` setting that selects the Ollama model the knowledge base is embedded and searched with. EMBEDDING_MODELS lists each supported model with its vector width, task prefixes and a calibrated default relevance floor: nomic-embed-text:v1.5 768 default, unchanged bge-m3 1024 multilingual qwen3-embedding:0.6b 1024 multilingual snowflake-arctic-embed2 1024 multilingual embeddinggemma 768 multilingual An unset setting behaves exactly as before. Score scales differ between models, so the relevance floor is now per model: an unset `rag.minRelevance` resolves to the configured model's floor, and the eval harness pins to it. Each floor was calibrated the way RAG_MIN_FINAL_SCORE was, with `eval:retrieval --min-final-score` sweeps. Under nomic's 0.62, bge-m3 loses a tenth of the golden set. Switching models is followed by the existing Reset & Rebuild. The model a collection was built with is recorded when it is created (collections from before this change count as nomic), and until they match embed jobs fail once with UnrecoverableError instead of retrying 30 times, and search returns no context with a warning. This also catches switches between models of the same width, which Qdrant would accept silently. Embedding models whose names lack "embed" (bge-m3) are also kept out of the chat model lists and recognised by GPU pacing, the configured model is exempt from the VRAM unload sweep, and only nomic keeps the any-tag installed-model fallback, since other tags of a family can differ in width (qwen3-embedding:8b is 4096-d). |
||
|---|---|---|
| .. | ||
| eval | ||
| functional | ||
| standalone | ||
| unit | ||
| bootstrap.ts | ||