Retrieval quality is usually a content problem, not a model problem
When an assistant answers badly, the source library is more often at fault than the model behind it.
By Xonique Editorial TeamEditorial Desk
Published · 6 min read
A recurring pattern in internal assistant projects: the team upgrades the model, sees a marginal improvement, and concludes the technology is not ready. An audit of the underlying document set usually tells a different story.
Three content failures that look like model failures
- Two documents state contradictory policies and neither is marked authoritative.
- An outdated procedure was never removed from the shared drive.
- The same guidance exists in six near-identical copies with small differences.
You cannot retrieve your way out of a knowledge base nobody maintains.
What to check before you commit
- Audit sources before changing models.
- Mark one document per topic as authoritative.
- Remove superseded material rather than archiving it in place.
- Re-run your evaluation set after every content clean-up.
A note on measurement
Teams that treat retrieval quality as an engineering project usually measure the wrong thing. Instrument the business outcome first — cycle time, cost per transaction, resolution rate, revenue retention — then work backwards to the technical metrics that move it.
- retrieval
- knowledge management
- content
- ai
- llm