How to Build Source-Aware AI Answers

Provenance begins at ingestion

Keep connector identity, external document ID, canonical path, title, modification time and a content fingerprint. When content is split, every retrieval unit should inherit parent identity and enough location information to reconstruct its origin.

Do not use a mutable title as the only identifier. Titles collide and change; stable IDs let the interface show a human label while the system verifies the artifact.

Return evidence as data

Pair text with provenance instead of flattening everything into a prompt string. The answer can cite a document ID, and the application can render its trusted title and link without accepting generated URLs.

{ "document_id": "doc_…", "title": "Security handbook", "location": "Access control / page 14", "text": "…" }

Validate, then render

1. Require claims to reference evidence IDs.
2. Reject IDs absent from retrieval.
3. Check that the passage supports the claim.
4. Render links from trusted application metadata.
5. Show uncertainty when evidence conflicts or is incomplete.

A source can be real but irrelevant; a claim can be helpful but unsupported.

Design for verification

Give users a title, excerpt and source location. For sensitive workflows, record the retrieval query, source versions and evidence IDs used.

SPYN3 search returns document and connector identity, relevance score and snippet. See the source-aware response guide for current fields.