Systems integration
Stop tuning prompts. Start cleaning data
Prompt tuning has a low ceiling because it cannot fix what a retrieval system is actually reading. On one project, cleaning the knowledge base moved the hallucination rate further than three weeks of prompt engineering had.
Executive summary
3 weeks
of prompt engineering that moved the hallucination rate less than cleaning the knowledge base did, on one project
Core conclusions
- A retrieval-augmented system is bounded by the quality of the documents it retrieves — no prompt instruction makes a stale or contradictory source say something true.
- The failure is easy to misdiagnose as a model problem because the output is fluent and wrong, so attention goes to the generator rather than the source.
- High retrieval recall over a poor corpus still collapses — deduplicate, date everything, and remove superseded material rather than annotating it.
A retrieval-augmented system is a reading system. Its answers are bounded by the quality of the documents it retrieves, and no instruction in the prompt can make a stale, duplicated or contradictory source say something true.
The failure is easy to misdiagnose because it presents as a model problem. The output is fluent and wrong, so attention goes to the thing generating the text rather than the thing supplying it.
Data habits that move the number
- Deduplicate ruthlessly. Two near-identical documents with one conflicting figure will surface the wrong one roughly half the time.
- Date everything, and let retrieval prefer current over merely relevant.
- Remove superseded material rather than leaving it in with a note, which retrieval cannot read as an instruction.
- Check retrieval quality separately from answer quality. High recall over a poor corpus still collapses.
High retrieval recall does not rescue a weak corpus. If the documents themselves are low quality, accuracy falls regardless of how well the retriever performs — which is why this is an integration problem before it is an AI one.
