For business
What it is
The organised, versioned store of everything the system knows, filterable by client, date and document type.
Why it matters
Answers can be scoped (only this client, only this year) and reproduced later: you can show which version of the knowledge produced a report.
Risk without it
Stale or deleted documents keep appearing in answers, and nobody can reproduce last month’s output.
For engineers
How it works
Documents and chunks are stored with versions and lifecycle state; deletions propagate to every index (right to erasure). When chunking or embedding strategy changes, a background job builds a new index and traffic switches over once it is validated.
Typical stack
- S3
- Postgres / DynamoDB metadata
- OpenSearch / pgvector
- background jobs
Pitfalls & trade-offs
- Deletion must reach every copy, caches included.
- Blue-green re-indexing avoids downtime.
- Keep the knowledge version in every answer’s citations.
In this example
In
→ index update
Out
kb v13 18 docs for A-17
Fictional sample · not a client system