What the documents look like
Retail checkout and loyalty work depends on four document families: promotional price tags, membership points rules, daily checkout reports, and exception refund slips. These materials come from the POS system, the membership CRM, the payment platform, and the price tag system, and they update daily or once per shift. A price or membership benefit error can produce price complaints, cash shortfalls or overages, and disputes over member benefits. Product information and selection questions therefore ask where a specific promotional price, points rule, or refund appears, and which document supports it. Citation tracing has to locate the exact document and version behind each cited statement.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model. Cited content is output as an array, and its length may be zero; when nothing is found, the output path still runs. The citation cap is expressed in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and that threshold is a filter condition. It does not carry permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF, reranking is enabled according to settings, and image retrieval requires image input plus the corresponding model capability. Checking whether a document is current requires the current and historical versions to exist side by side, and whether a citation is adopted still depends on a valid-version review.
Hypothetical drill sample: in store handover materials, a promotional price tag document states "June 1, 2024 to June 30, 2024, product A promotional price 9.9 yuan." A membership points rule states "during the activity period, points for product A are calculated on the actual amount paid." A draft handover request states "product A continues at the promotional price of 9.9 yuan on July 1, and points accrue as usual." The request date falls outside the effective range stated in the price tag document, and the points rule does not state whether the activity continues into July.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point or value method | Comparison and decision method |
|---|---|---|
| Associated knowledge bases | Select knowledge bases sharing the same vector model | Check the vector model identifier of each selected base; exclude any base whose identifier differs |
| Search mode | Hypothetical starting point for this drill is hybrid retrieval | Compare recall behavior against semantic retrieval and full-text retrieval separately |
| Minimum relevance | Hypothetical starting point for this drill is 0.5 | Compare the change in result count against no threshold or a higher threshold |
| Citation cap | Hypothetical starting point for this drill is 2000 tokens | Compare citation count against a smaller or larger token cap |
| Result reranking | Hypothetical starting point for this drill is enabled | Compare candidate ranking positions against reranking disabled |
| Document currency check | Prepare both current and historical versions | Compare the version and date range of each cited passage against the request |
Three common errors
- Treating the June price tag document as the basis for the July 1 handover request. The price tag text states only June 1 to June 30; whether the price applies on July 1 is unsupported and requires separate verification.
- Treating minimum relevance as a permission filter. That threshold only filters low-relevance results; it cannot exclude unrelated business lines or unauthorized records, and retrieval itself does not constitute authorization.
- Treating an empty citation array as an error. The documentation states the citation length may be zero and the path still runs; whether an error occurs requires separate observation.
How to verify the setup
- Input the price tag text and the July 1 draft request from the hypothetical sample, compare the date range in the cited content against the request date, and observe whether the system returns the June 1 to June 30 passage. If the returned content lacks a date range or returns a different period, the condition is not met.
- Input the same draft request and compare whether the answer directly confirms July execution. If the answer contains a definite execution statement, the condition exceeds what the source text supports.
- Input the paper value
[]and check the downstream handling design, comparing whether the downstream node treats it as an empty citation. If the downstream node treats the empty array as an error and stops, the handling design is not met. All three checks are paper reviews and were not actually run.