Source Provenance in RAG
Track every answer back to its source documents with chunk-level attribution, version history, and audit trails that meet enterprise compliance requirements.
When a chief of staff asks Donna for the latest parental-leave policy, the answer must cite the exact paragraph in the employee handbook — not because it looks professional, but because HR needs to verify the claim before publishing it to the team. Source provenance is the mechanism that connects every generated sentence to the document chunk, version timestamp, and retrieval confidence score that produced it.
OpenBase builds provenance into the retrieval layer. Each chunk carries metadata: document title, author, last-modified date, section heading, access permissions. When the LLM synthesizes an answer, the system maps each claim back to the chunks that supported it, ranks them by retrieval confidence, and surfaces citations inline. The reader clicks a citation and lands on the exact paragraph in the original document. If the source document updates, the system flags answers that reference the old version.
This is not decoration. Regulated industries require audit trails showing which documents informed which decisions. Internal comms teams need to verify claims before they reach 500 employees. Legal needs to prove that an AI-generated contract clause came from an approved template, not a hallucination. Provenance makes the answer verifiable; without it, the answer is a liability.
Problem & solution
Unverifiable AI answers erode trust
A marketing lead asks the AI workspace for the brand voice guidelines. The answer sounds plausible, but there is no citation. She cannot verify whether it came from the official brand book or a draft someone uploaded last year. She rewrites the answer manually, and the AI workspace becomes a suggestion tool instead of a decision tool. Without provenance, every answer requires human re-verification, and adoption stalls.
Chunk-level attribution with audit trails
OpenBase tracks provenance at the chunk level. When a document enters the knowledge base, the ingestion pipeline tags each chunk with document metadata, section hierarchy, and version timestamp. During retrieval, the system scores each chunk by semantic relevance and recency, then maps the LLM's answer back to the top-ranked chunks. The UI displays citations inline, linking directly to the source paragraph. If the source document updates, the system flags dependent answers and re-indexes the affected chunks. Every answer includes a provenance footer showing which documents contributed, their last-modified dates, and the retrieval confidence scores.
What you see after 90 days
- Compliance-ready audit trails linking every answer to source documents and retrieval timestamps
- Inline citations that land the reader on the exact paragraph that supported the claim
- Automatic flagging of answers when their source documents update or become inaccessible
- Retrieval confidence scores that let the reader judge evidence strength before acting on the answer
- Version history showing how an answer changed as the underlying knowledge base evolved
Who benefits most
- Legal and compliance teams that must prove which documents informed which decisions
- Internal comms leads who verify AI-generated announcements before publishing to the company
- HR teams citing policy documents in answers to employee questions
- Finance controllers referencing specific contract clauses or approval workflows
- Ops managers who need to trace process changes back to the updated SOP
Frequently asked questions
What happens when a cited source document is deleted or moved?
OpenBase flags all answers that reference the deleted document and marks the citations as broken. The system does not remove the answer, because the historical record matters for audit purposes, but it surfaces a warning to the reader and prevents the answer from appearing in new searches until a human reviews it.
How granular is the citation? Does it link to the document or the specific paragraph?
Citations link to the chunk level — typically a paragraph or section heading. The reader clicks the citation and lands on the exact text that supported the claim. If the document is a PDF, the link includes a page number and text anchor. If it is a wiki page, the link includes a section fragment.
Can I see which version of a document was cited when the answer was generated?
Yes. The provenance footer includes the document version timestamp. If the document has been updated since the answer was generated, the system displays both the original version cited and a link to the current version, so the reader can compare them.
How does provenance work when the answer synthesizes information from multiple documents?
The system attributes each claim to its supporting chunks and displays multiple citations inline. If two documents conflict, the system surfaces both and flags the conflict with a note. The reader sees which document was retrieved with higher confidence and can decide which source to trust.
Does tracking provenance slow down answer generation?
Provenance metadata is retrieved in the same query that fetches the chunks, so the latency cost is negligible — typically under 50ms. The system pre-indexes metadata during document ingestion, not at query time.
What metadata does OpenBase track for each source document?
Document title, author, last-modified timestamp, section hierarchy, access permissions, file type, and a unique document ID. Custom metadata fields can be added during ingestion if your compliance requirements demand additional tracking.
How do I audit which documents contributed to answers over the last quarter?
The audit log exports a CSV showing every answer generated, the documents cited, the retrieval confidence scores, and the timestamps. You can filter by document, by user, or by date range. This is the report compliance teams use to prove which knowledge base informed which decisions.
Can I hide provenance from end users but keep it for audit purposes?
Yes. The tenant admin can configure whether citations appear inline in the answer or only in the audit log. Some teams show citations to all users; others restrict them to admins and show end users only a summary confidence score.
In this cluster
Hub: enterprise-rag-architecture