STEP 01
Make each answer self-contained
An isolated sentence such as “It supports two devices” is ambiguous when several products share an index. Include the product and question in the same passage.
- Keep each question with its answer, product name, version, and important limitations.
- For tables, preserve the column headers and units with each row. Inspect extracted text instead of assuming the PDF layout survived.
- For sums, inventory counts, or joins, use a structured query workflow. This text-retrieval recipe does not validate calculations over tables.
STEP 02
Test wording and meaning
Use Hybrid Search as a baseline and try a keyword-heavier balance for identifiers. Treat suggested weights as experiments, not measured recommendations for your collection.
- Try the exact product code, a partial description, and a natural-language question about the same item.
- Compare the returned product and evidence before changing the answer model.
- When multiple versions coexist, add version metadata and use an appropriate filter. Ask for clarification if the intended version is unknown.
STEP 03
Use a discriminating test pair
Example test design: ask the same compatibility question about two different products whose answers differ. Substitute products and verified answers from your own source documents.
- Record which document and version should support each answer.
- Pass only if both the product identity and its limitation are preserved.
- Add an unknown product code. The assistant should request clarification or say the evidence is missing rather than borrowing another product's answer.
Give your changes a fair test.
Open comparison worksheet Use the same questions, inspect the evidence, and record what changed.