CORE GUIDE

SYSTEM_COMPONENTINTERMEDIATE7 min read

Knowledge base lifecycle

A useful AI knowledge base is a maintained evidence system with ingestion, authority, update, retirement and evaluation policies.

Mental model

Knowledge is not 'uploaded' once. Every source has an owner, effective period, authority level, ingestion state and retirement path that determine whether it should still influence answers.

Why it matters

RAG systems decay when old policies remain searchable, duplicates accumulate and nobody knows which source supersedes another. Retrieval tuning cannot fix a corpus whose lifecycle is undefined. Treating the knowledge base as an operational system makes freshness, authority and deletion explicit responsibilities rather than occasional cleanup.

01

Manage sources from admission to retirement

Define who may add sources, required metadata, parsing and chunking rules, authority and effective dates, update triggers and expiry behavior. Track source versions and supersession so retrieval can prefer controlling material. Re-index changes safely, run retrieval regression tests, and retain enough provenance to explain which source supported an answer.

02

Example: refund policy changes from 30 to 14 days

The new policy is ingested with an effective date and marked as superseding the previous document. The old policy is retained for historical audit but excluded from current-support retrieval. A regression query verifies that the 14-day source is now retrieved and the stale policy no longer controls the answer.

Common failure modes

  • Leaving superseded documents searchable with no authority metadata.
  • Treating ingestion success as proof the content is retrievable for real questions.
  • Deleting or replacing sources without retaining version and provenance history.

Engineering heuristics

  • Assign ownership and update triggers to every authoritative source class.
  • Represent supersession and effective dates explicitly.
  • Run retrieval regression checks after meaningful corpus changes.

Takeaways

  1. 01Knowledge quality depends on lifecycle, not just embeddings.
  2. 02Freshness and authority need operational owners.
  3. 03A living knowledge base must support update, retirement and evaluation.

Related concepts from the Knowledge Graph

These relationships come from the canonical graph, not a separate Guide taxonomy.

Customer-support copilot and escalationENABLES