Retrieve Knowledge Base

Embed query and return the top-k matching chunks from the KB.

Auth: write:agents. Chunks are customer data, same threat model as /agents/{id}/logs.

Tenant scoping: a kb_id from another tenant returns 404 just like the read endpoint — no existence leak.

Rate limit: BUCKET_KB_RETRIEVAL at 30 req/min/tenant — same budget as the other outbound-call endpoints (MCP discovery, validate-llm, models-live).

Dual-surface model: this endpoint is one of three KB retrieval paths. The Java agent runtime queries Pinecone directly at deploy time via its own mirror (off the hot path through this BE); the streamkap-tools MCP server exposes streamkap_kb_retrieve and proxies through this endpoint for third-party agents.