Knowledge Base

The knowledge base gives your agent access to company-specific information. It uses a hybrid retrieval system combining vector similarity, keyword matching, and conversation memory.

Adding documents

Upload documents via POST /api/knowledge:

curl -X POST https://your-deployment.convex.site/api/knowledge \
  -H 'Authorization: Bearer SESSION_TOKEN' \
  -d '{
    "title": "Return Policy",
    "content": "Items can be returned within 30 days...",
    "category": "policies"
  }'

Documents are automatically chunked and embedded for vector search.

Searching

Use POST /api/knowledge/search for hybrid retrieval:

• Vector search — Semantic similarity using embeddings (weight: alphaVector, default 0.62)
• BM25 keyword search — Traditional keyword matching (weight: alphaKeyword, default 0.38)
• Transcript memory — Recent conversation context boosts relevance (weight: alphaMemory, default 0.20)

Knowledge gaps

Metrics

Track search quality with GET /api/knowledge/metrics:

• Total searches — How often the knowledge base is queried
• Average top score — Quality of the best match
• Gap rate — Percentage of searches that produce no good results