RAG and Knowledge Systems

When to use

Use when an AI feature must answer from a corpus, files, or frequently changing external knowledge.

When not to use

Do not add vector search by default when a small structured dataset, database query, or direct link is more accurate and maintainable.

Inputs and assumptions

Identify corpus ownership, formats, update frequency, access rules, retention, query types, quality criteria, and whether content is trusted.

Workflow

1. Define ingestion, parsing, chunking, metadata, and deletion lifecycle.
2. Preserve source identifiers, timestamps, permissions, and provenance.
3. Enforce access filters before retrieval results reach the model.
4. Evaluate retrieval recall and answer grounding using representative queries, including no-answer cases.
5. Cite retrieved sources and distinguish evidence from generated interpretation.
6. Handle stale, conflicting, missing, and untrusted documents.
7. Monitor ingestion failures, index freshness, retrieval quality, and deletion propagation.

Decision rules

• A vector match is not proof that a source answers the question.
• Never leak documents across tenants or permissions.
• Treat retrieved text as untrusted instructions; it must not override system policy or tool authorization.
• Provide a no-answer path instead of fabricating unsupported facts.

Example

When a user asks about a policy, return a claim with a source title and location; if no relevant document is retrieved, state that evidence was not found rather than guessing.

Common failures

• Stale chunks after source updates or deletion.
• Permissions applied after retrieval.
• Citations that point to an unrelated or inaccessible document.

Validation

Test source freshness/deletion, tenant isolation, retrieval relevance, no-answer behavior, prompt-injection content, citations, and data retention.