# Build internal knowledge answers with a retrieval test

> Prepare an approved document corpus and test citations, access boundaries, and unanswerable questions before opening an internal assistant.

- Canonical: https://gptnavi.com/workflows/internal-knowledge-answer-with-retrieval-test
- Category: Operations
- Difficulty: Advanced
- Setup time: 4-8 hours
- Estimated time saved: Measure in your pilot; no guaranteed saving
- Last materially updated: 2026-09-06
- Best for: Internal knowledge teams
- Tools: Notion, Docling, Dust, Google Sheets, Stack AI, Jina AI, Voyage AI

## Quick answer

There are two implementation routes: a managed Dust or Stack AI assistant, or a developer-built retrieval pipeline. The parser and embedding providers are optional components for the latter, not guaranteed native integrations with every managed platform.

## When to use it

- Product policy answers
- Operational handbook search

## Steps

1. **Approve a small corpus** — Select twenty current documents with owners and access groups. Exclude secrets and expired policies. Write fifteen answerable questions and five questions that must be refused or escalated. Tool: Notion. Expected output: Versioned corpus and test questions
2. **Inspect extraction** — For a custom pipeline, run Docling locally and inspect tables and page order. Use Jina AI Reader only for approved web documents where cloud processing is allowed. Keep source IDs and access metadata attached to chunks. Tool: Docling. Expected output: Audited text chunks with provenance
3. **Choose a deployment route** — For a managed pilot, connect approved sources in Dust or Stack AI and test permissions. For a custom pipeline, benchmark Voyage AI against Jina embeddings or rerankers on your labelled questions; write the integration explicitly rather than assuming compatibility. Tool: Dust. Expected output: One restricted pilot assistant
4. **Gate access on evidence** — Run the twenty questions as two different permission groups. Check citations against exact source passages, mark unsupported answers wrong, and require zero cross-group disclosures. Keep an escalation path for missing answers. Tool: Google Sheets. Expected output: Acceptance report with failures and owners

## Prompt templates

### Adaptable task prompt

Answer [question] using only the supplied authorized passages [passages]. Cite the exact document and section. If evidence is missing or conflicting, state that and request review. Do not expose documents outside the requester's access group.

## Common mistakes

- Assuming managed tools accept every external embedding provider
- Checking answer fluency instead of evidence and permissions

## Related workflows

- https://gptnavi.com/workflows/inbound-request-to-approved-owner-task