What the documents look like
Convenience-store fresh food is ordered by the store at high frequency against a short shelf life, with no joint-operation revenue split of the kind used at hypermarket fresh counters. The working documents for this category are the fresh-food order form, the loss report, the near-expiry handling sheet, and the temperature record. Data comes from POS sales, the headquarters ordering system, store inventory, and cold-chain temperature logs.
For copy drafting, this matters because promotional sentences about shelf life, temperature status, or loss quantities must be traced back line by line to those documents. A sentence that reads well can still be unsupported by the material. Fields the documents do not carry remain unknown, and the fact that other categories handle more online certification or electronic contracts does not support any conclusion about this category.
Constraints at the citation tracing step
The knowledge base search node can link one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. The citation cap is measured in tokens. The minimum relevance value takes effect only under semantic search or result reranking. Retrieval does not perform authorization checks, and a similarity score is not a category, entity, or permission filter. Hybrid search merges results through RRF. Reranking requires a rerank model and must be enabled by configuration. Image retrieval requires image input and the relevant model capability. Optional paths in the overall retrieval flow do not mean every request executes all of them. Knowledge bases built on different vector models cannot be linked together; the identifier is checked and non-matching bases are excluded.
Hypothetical paper exercise. The material text is: "On 10 March 2025, the fresh-food order form for Store A recorded: rice balls, 40 units ordered; arrival temperature log, 2°C. On the same date, the near-expiry handling sheet recorded: sandwiches, 8 units unsold before 20:00, transferred to loss report." The draft sentence to be judged is: "Fresh food at this store is cold-chain throughout, and anything unsold on the day is reported as loss, so expired goods never appear." The words "throughout" and "never" have no corresponding record in the material. The material covers only two items at Store A on 10 March 2025 and cannot support a conclusion about the whole store across all periods.
Configuration and implementation checks
| Check item | Hypothetical starting point or setting | Comparison method |
|---|---|---|
| Linked knowledge bases | Select bases containing order forms and temperature records | Compare the vector model identifier of each base; exclude any base whose identifier does not match |
| Citation cap | 800 tokens as the starting point for this exercise, because the sample text is short | Compare against the actual token count of the sample text and observe whether truncation occurs |
| Minimum relevance | 0.4 as the starting point for this exercise | Compare against the semantic search or rerank score and observe whether the sample is filtered out |
| Search mode | Hybrid search, because the sample contains a date and item names | Compare against full-text search alone and observe differences in matching identifier-like strings |
| Citation output | Carries the citation array | Use the paper input [] to check whether the downstream path still completes |
| Citation display check | Compare the answer against the source text line by line | Check whether each clause of the draft sentence has a corresponding field in the source |
Three common errors
- Treating minimum relevance as a permission filter. The input is the sample question and the knowledge base configuration; the comparison is whether the value acts only on semantic search or rerank scores; the observable result is that unrelated business records may still be recalled.
- Treating retrieval output as a verified source. The input is the citation array and the draft sentence; the comparison is whether the phrase "cold-chain throughout" has a corresponding temperature record; the observable result is whether a conclusion drawn from single-day, single-store material exceeds that scope.
- Treating an unrecorded field as a denial. The input is the source text and the draft sentence; the comparison is whether "expired goods never appear" has a corresponding entry in the loss report or near-expiry handling sheet; the observable result is that a missing record is logged as unsupported.
How to verify the setup
- Input the sample source text and the draft sentence. Compare whether each clause of the draft sentence matches the source on subject, date, item name, and status. Observe that when "throughout" and "never" have no corresponding record, the clause is judged unsupported.
- Input the paper citation array
[]. Compare whether the downstream answer step still completes. Observe that the empty array does not cause the workflow to error out or terminate, which meets the design expectation. - Input two knowledge bases with different vector model identifiers. Compare the linking condition. Observe that the non-matching base is excluded and no cross-model retrieval is executed, which meets the check condition.