What the documents look like
Fresh fruit store procurement draws on four source types: wholesale market price information, supplier quotations, store sales, and inspection reports. The daily documents are purchase orders, acceptance records, pesticide residue inspection reports, and shrinkage records. Buyers, store fresh-produce supervisors, and wholesale market stalls all take part, and materials update daily or per batch.
Citation tracing for this category answers one question: can a price, grade, or inspection conclusion that appears in an answer be traced back to the wording of the corresponding document? Market prices and quotations change by batch, and the same supplier may deliver different grades and prices across batches. Tracing therefore requires checking document subject, batch, and date range together. A single quotation cannot be read as a long-term supply condition.
A knowledge-base search node outputs an array of cited content, and that array may have length zero. An empty array still completes the output path. Linked knowledge bases must share the same vector model; bases with different vector models cannot be linked together. The minimum relevance value is a filter threshold that applies only in semantic retrieval or when result reranking is used, and that value carries no permission or entity-isolation role. The citation limit is measured in tokens and is a different field from a thinking budget; the two cannot substitute for each other. Hybrid retrieval merges multiple result paths through RRF, reranking requires a configured model and must be enabled by settings, and image retrieval requires image input plus matching model capability. These paths do not all execute on every request. Whether a retrieved result is adopted still depends on human review against a valid version.
Hypothetical paper exercise sample: a store acceptance record dated 12 March states "Red Fuji apple 80# one carton, 20 kg, acceptance passed." A pesticide residue inspection report of the same day states "the sample from this batch meets the limit requirements." A supplier quotation states "this quotation is valid until 15 March." The draft sentence to be judged is "this supplier's apple supply price is stable and valid long term for the current quarter." The retrieval scope is limited to those three materials. The authorization boundary is internal store verification, with no external commitment.
Constraints at the citation tracing step
The search node returns cited content as an array whose length may be zero, so a downstream branch must be designed to accept an empty array rather than treating it as a fault. Linking knowledge bases requires identical vector models; a base whose model identifier differs is excluded before any search runs, and no cross-model search is performed to compare recall.
The minimum relevance value filters low-relevance results and takes effect only under semantic retrieval or result reranking. It does not exclude unauthorized records or unrelated businesses. Retrieval itself is not authorization. Any permission boundary must be implemented downstream.
The citation limit is expressed in tokens, not in a count of items. A thinking budget is a separate field and cannot be repurposed as a citation limit. RRF is a fusion process applied to hybrid results; it is not a switch that a user enables. Reranking and image retrieval are optional paths that run only when a model is configured, the setting is enabled, and the required input exists.
Judgment must hold subject, period, and field constant. Contradiction applies only when the same field carries explicitly exclusive values. A value that is not stated, or that is supported only by a related stage, counts as unsupported or pending verification. A quotation validity date is not a supply term. Pending review does not show that an operation right is absent. Signing, reaching cooperation, and contract effectiveness are three distinct facts.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select only procurement material bases sharing one vector model | Compare vector model identifiers across bases; exclude any base whose identifier differs |
| Search mode | Hybrid retrieval | Compare against semantic retrieval and full-text retrieval; grade and numbered queries need keyword hits |
| Minimum relevance | Hypothetical starting point 0.4 | Compare returned item counts at 0.3 and 0.5, and observe whether unrelated fragments decrease |
| Citation limit | Hypothetical starting point 1500 tokens | Compare against 1000 and 2000, and observe whether citations cover the key fields of the documents |
| Result reranking | Enabled | Compare candidate ordering against the disabled state, and confirm a rerank model is configured |
| Response format | Default | Compare against json_schema; this task has no fixed-structure reading requirement |
The values above are hypothetical starting points for this paper exercise, not official default values.
Three common errors
- Treating a quotation validity date as a supply term. Input is the quotation text; the comparison object is the acceptance record date. The quotation is valid until 15 March and acceptance occurred on 12 March. These are different fields, and the draft sentence's "valid long term" is unsupported.
- Treating an empty citation array as a workflow fault. Input is the retrieval output
[]; the comparison object is the node output description, which permits length zero. The result is recorded as a downstream design item to check, not as a node error. - Treating minimum relevance as a permission filter. Input is the 0.4 threshold; the comparison object is whether unauthorized records are excluded. The value acts only on relevance, permission requires a downstream implementation, and the claim is unsupported.
How to verify the setup
- Input the three source texts above together with the draft sentence. Compare "current quarter," "stable," and "valid long term" in the draft against the subject, period, and field stated in each source. A finding of support requires the source to state the corresponding value explicitly. In this exercise the finding is unsupported.
- Input the retrieval output
[]together with the downstream node configuration. Compare whether the downstream branch accepts an empty array. If the downstream continues without error, the acceptance design holds. - Input the vector model identifiers of the selected knowledge bases. Compare whether the identifiers match. Bases with differing identifiers are excluded, and the observed result is recorded. No cross-model search is run to compare missed recall.