Local evidence embeddings

Related-evidence lookup

The updated agent can use local EmbeddingGemma vectors to find related passages in the current question’s SQLite evidence cache. Exact counts, indicator matching and complete inventory still use SQL and the existing chunk-review path.

Enable on an endpoint

Install embeddinggemma in local Ollama if needed, then merge these fields into the agent configuration and restart/reload that agent with the updated binary.

{
  "embeddings_enabled": true,
  "embedding_url": "http://127.0.0.1:11434/api/embed"
}

Embedding inference is loopback-only and has no hosted-provider credential. The answering model is configured separately; retrieved text may still be sent to that model if it is hosted.

Retrieval boundaries

• One private in-memory SQLite vector cache per question; no cross-customer or cross-conversation index.
• Up to 1,024 chunks, 1,000 characters each with overlap; source and candidate limits stay visible.
• Background indexing with deadlines; text matching if embeddings are pending or unavailable.
• Pack reloads invalidate stale vectors. Results retain source pack, row, chunk and cache timestamp.
• Ranking scores are not confidence. A missing match never proves absence.