What the documents look like
Cold-chain receiving at a store turns on transport temperature, arrival batch, and the store's receiving outcome. It does not cover origin pricing. The working materials are cold-chain temperature records, arrival acceptance forms, rejection forms, and supplier delivery notes. They come from reefer temperature control, the WMS, the store receiving system, and quality-inspection photographs. Data updates per batch and per day. The three most time-consuming steps are measuring temperature, sampling quality, and registering rejection reasons. An operating report or exception review has to answer which material, which batch, and which handover step a temperature reading or acceptance conclusion came from. When a citation is separated from its source text, responsibility for spoiled food and disputes with suppliers become hard to pin down.
For this page the question is narrow: at the citation tracing step, how does a reviewer connect a statement in a report back to the underlying receiving document.
Constraints at the citation tracing step
A knowledge-base search node can associate one or more knowledge bases that use the same vector model. Its output is an array of cited content, and that array can have length zero. When no content is found, the output path still runs through. Retrieval results are handed to the answering step, where each answer is checked line by line against the cited source text. A reasoning trace or a fluent answer does not substitute for source evidence. Document timeliness has to be tested with both the current version and a historical version of a material; whether a citation is adopted still depends on review of the effective version. The minimum relevance value is a filter threshold and does not act as a permission or entity separator. Retrieval itself is not authorization. Hybrid retrieval fuses multiple result paths through RRF, reranking requires the model to be present and enabled by configuration, and image retrieval requires a corresponding image input and model capability. Knowledge bases built on different vector models cannot be associated together; check the identifiers first and exclude those that do not match.
Hypothetical drill sample: a store receiving team keeps one cold-chain temperature record whose original text reads "12 March 2025, batch LJ-0312, cold cabinet temperature recorded at 6.5°C, signed and confirmed by the receiving clerk; the arrival acceptance form for the same batch notes 'temperature out of range, pending quality-inspection review'." A draft sentence reads "This batch passed temperature check and has completed acceptance." To be judged: whether the draft sentence agrees with the same-field values in the two materials, and whether "pending quality-inspection review" constitutes completed acceptance.
Configuration and implementation checks
| Configuration or check item | Use on this page | Basis and boundary |
|---|---|---|
| Associated knowledge bases | Limit the material scope searched for this review | Only knowledge bases with the same vector model can be associated; verify identifiers first |
| Search mode | Batch and number lookups go through full-text or hybrid retrieval | The documentation describes three modes and does not name a default |
| Citation upper limit | Control the amount of tokens cited per search | Documented in tokens; top k is not used |
| Minimum relevance | Filter low-relevance results | Takes effect only in semantic retrieval or with reranking; not a permission filter |
| Cited content output | Carry retrieval results into source-text checking | Array format, length may be zero; an empty array does not mean the flow errored |
| Material timeliness check | Test sources against current and historical versions | Whether a citation is adopted still depends on effective-version review |
Three common errors
- Treating "pending quality-inspection review" as completed acceptance. Input is the draft sentence and the two materials; the comparison is the acceptance-status field for the same batch; the observable result is that the materials state no review conclusion, so the draft sentence goes beyond what the source text supports.
- Treating minimum relevance as a permission or entity filter. Input is the similarity threshold setting; the comparison is the documented conditions under which minimum relevance takes effect; the observable result is that the value filters low-relevance results only in semantic retrieval or with reranking, and does not exclude unrelated businesses or unauthorized records.
- Associating knowledge bases that use different vector models. Input is the vector-model identifiers of the several bases; the comparison is the documented requirement that the vector models be the same; the observable result is that bases with mismatched identifiers do not meet the association condition, and the reviewer records and excludes them without inferring run outcomes.
How to verify the setup
- Test with the two source materials and the draft sentence from the hypothetical sample. Input is the source text of the materials and the draft sentence; compare the temperature field and the acceptance-status field for the same batch; the observable phenomenon is that when the draft sentence disagrees with the source values and the review conclusion is absent, the sentence is judged unsupported and the evidence gap is preserved.
- Check empty-citation handling with the explicit paper input
[]. Input is a citation array of length zero; compare whether the downstream answering step still executes; the observable phenomenon is that the path runs through and no error assertion is produced, which matches the documented behavior for an empty array output. - Test source timeliness with a current version and a historical version of the material. Input is both versions for the same batch; compare the version identifier the citation points to against its effective scope; the observable phenomenon is that the citation can be traced to a specific version, while adoption still awaits effective-version review, which satisfies the source-checking requirement.