Separate Model Errors from System Errors

A practical approach

Step 1

Define the expected input and output of each stage. Record identifiers and safe diagnostic context across the request.

Step 2

For a failed case, inspect the source, parsed data, retrieved evidence, tool results and final rendering in sequence.

Step 3

Reproduce the failing stage with a controlled fixture and add a targeted check. Re-evaluate the complete journey after the fix.

Worked example

An assistant quotes an outdated course description because the ingestion job never removed the previous document version. The model may faithfully summarize what it received. Fixing index replacement addresses the system error more directly than repeatedly changing the answer prompt.

A mistake to avoid

Blaming the model for every issue can hide ordinary software defects. The opposite mistake is assuming correct tool output guarantees a correct answer; generation and presentation still need checks.

Try it yourself

Take one failed output and draw its data path. Mark the first point where the actual data differs from the expected contract, then design a minimal reproduction for that point.

Your implementation checklist

• Stages have explicit contracts.
• The first divergence is identified.
• Fixes target a demonstrated cause.
• The end-to-end result is rechecked.