Skip to main content
knowledge.search performs a governed similarity search over your organization’s uploaded knowledge base documents. The model receives the top-matching excerpts, ranked by similarity, with citations back to the source document. HASP owns the embedding pipeline end-to-end — chunking, embedding, and vector search all happen inside HASP’s infrastructure. Document RAG is off by default per org. An admin must upload documents and enable retrieval under Settings → Knowledge Base before knowledge.search is available to the model.

​
PHI handling

Unlike web.search, PHI handling for knowledge.search happens at two different points:
  1. At ingestion — every uploaded document is scanned and PHI is redacted before it is chunked and embedded. This is unconditional: there is no setting that disables it, and no re-identification path exists for retrieved chunks. Only the redacted text is ever stored or searched.
  2. At query time — the search query itself is scanned and redacted before it reaches the embedding provider, the same posture web.search applies to outbound queries.
Because ingested content is always redacted first, retrieved excerpts can never surface PHI, regardless of what a user searches for.

​
Billing

Document RAG has no dedicated per-call surcharge (unlike web.search’s flat credit fee). Cost is metered through your existing allotments:
  • Storage — cumulative bytes of uploaded documents and their embeddings.
  • AI Credits — the embedding call at ingestion time, and normal input-token cost for retrieved excerpts injected into context.

​
Input parameters

The number of chunks returned per query is capped by your org’s max_chunks_per_retrieval setting (Settings → Knowledge Base), not a per-call parameter.

​
Output

Matching chunks are assembled into the model’s context as [n] {document title} followed by the chunk’s redacted text, using the same [n] citation-marker convention as web.search. If nothing matches, the model receives a “no matching knowledge base content found” result.

​
Citations

Retrieved chunks are persisted as citations linked to the conversation message, distinguished from web citations by kind: "document". A document citation has no external URL — only the source document’s title and rank. Citations survive page refresh and are available in the Audit endpoint.

​
Uploading documents

Documents are uploaded through the Public API’s Knowledge Base endpoints, gated by the knowledge:read scope for listing and the ai:embeddings scope for uploads and deletes. Uploads are processed asynchronously: the endpoint returns immediately, and the document’s status field (pending → extracting → chunking → embedding → ready, or failed) reflects ingestion progress. Uploading the same file content twice for the same org is a no-op — HASP deduplicates by content hash and does not re-embed. Accepted formats: PDF, DOCX, TXT, Markdown, HTML, and common image formats (scanned pages are extracted via HASP’s self-hosted OCR sidecar — no third-party OCR provider is involved).

​
Error codes

See Errors for the full error envelope shape.