What the documents look like
Fruit-store procurement draws on wholesale-market price movements, supplier quotations, store sales and inspection reports. The working documents are purchase orders, acceptance records, pesticide-residue inspection reports and loss records, and their update rhythm is daily or per batch. In cross-border and multilingual retrieval, one batch of fruit may correspond to a Chinese purchase order, a foreign-language inspection report and a quotation in a third language. When versions and languages diverge, tracing a cited source back to a specific batch becomes difficult, and a citation must be checked against the version that matches the current batch.
The knowledge-base search node can be linked to one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length may be zero. That output still flows through the chain when nothing is retrieved. Retrieved content is then handled by the answering stage, where each answer sentence is checked against the cited source text. The minimum relevance score is a filter threshold that applies only in semantic search or when result reranking is used; it does not perform permission or entity isolation. RRF is the fusion step inside hybrid search, not a standalone switch. Image retrieval requires a corresponding image input and model capability, and reranking requires a rerank model enabled in the settings. Optional paths in the overall flow do not mean every request executes all of them.
Constraints at the citation tracing step
A traceable citation has to connect the same subject, the same period and the same field. Where a document does not state the linking field, or where only a related stage is documented, the claim remains unsupported or pending verification. A pending review does not establish that an operating permission is absent. Signing, reaching cooperation and a contract taking effect are three distinct facts. A quotation validity period does not state a supply period.
Hypothetical paper exercise sample. A store procures a batch of imported oranges. The materials are a Chinese purchase order (number CG-2024-0512, dated 12 May 2024, quantity 200 boxes, supplier being overseas orchard A) and a foreign-language pesticide-residue inspection report (report number LAB-2024-0477, dated 10 May 2024, conclusion stated as qualified). A draft sentence reads: "This batch of oranges has passed pesticide-residue inspection and can be scheduled for shelving." To be judged: the inspection report's subject is overseas orchard A, while the purchase order's subject is the store, so the subjects differ; the report date precedes the purchase order date, so the periods differ; and the report does not state the corresponding purchase order number. The draft sentence ties an inspection conclusion directly to the procurement batch without same-subject, same-period and same-field evidence, so it is unsupported and requires an inspection report for the corresponding batch or an explicit manual verification record.
A second check concerns the empty-reference branch. A paper input of an empty array [] can be used to confirm whether the downstream design handles an empty citation array. This does not assert that any particular query returns an empty array.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the purchase-order base and the inspection-report base that use the same vector model | Check the vector model identifier and exclude bases with a different model; do not run cross-model retrieval |
| Search mode | Hybrid search | Purchase-order number lookups need full-text hits while descriptive queries need vector recall; compare how each mode recalls the sample |
| Reference upper limit | Set to 2000 tokens | Hypothetical starting point for this exercise; compare against the total length of the sample materials so the citation covers key fields of the order and the report |
| Minimum relevance | Set to 0.3 | Hypothetical starting point for this exercise; compare against the scores of irrelevant fragments in the sample and observe whether low-relevance results are filtered |
| Question optimization | Enabled | The sample contains multilingual materials; observe whether coreference removal improves cross-language retrieval |
| Citation handoff | Check whether the output array is passed to the answering stage | Verify each answer sentence against the cited source text, and confirm the downstream design handles an empty array |
Three common errors
- Treating the minimum relevance score as a permission or entity-isolation mechanism. Input: the sample purchase order and report. Compare whether retrieval results include materials from unrelated suppliers. Observe whether unrelated records are still returned. A relevance score does not perform authorization filtering.
- Treating RRF as a standalone switch to configure. Input: a hybrid-search setting. Compare the semantic and full-text result paths. Observe whether the fused ordering changes. RRF is a fusion step, not an option that is turned on.
- Treating a pending inspection report as qualified and scheduling shelving. Input: the sample report and the draft sentence. Compare the report subject and date with the purchase order. Observe whether same-subject, same-period evidence exists. Pending review does not equal approval, and a qualified report conclusion does not automatically equal a qualified procurement batch.
How to verify the setup
- Input the sample purchase order and inspection report. Compare whether the citation array contains the source text of both materials. Observe whether each answer sentence corresponds to a cited source. If the citation array is empty, check whether the downstream design handles an empty array; do not claim that retrieval necessarily returns empty.
- Input the draft sentence "This batch of oranges has passed pesticide-residue inspection and can be scheduled for shelving." Compare the report subject with the purchase order subject and the report date with the purchase order date. Observe whether same-subject, same-period, same-field evidence exists. When the subjects differ or the periods differ, judge the sentence unsupported and record the evidence gap.
- Input the purchase-order base and the inspection-report base that share the same vector model. Compare the vector model identifiers. Observe whether they can be linked together. When the identifiers differ, exclude that base and record the actual observation; do not run cross-model retrieval to compare missed recall.