What the documents look like
Wholesale market stall trading is more fragmented than regional distributor replenishment. It relies on on-site bargaining and shipping, and the main data sources are market price sheets, stall inventory, WeChat receipts, and logistics shipping notes. The documents handled day to day include purchase slips, handwritten sales orders, stall lease contracts, and payment receipts, with updates at daily or per-order frequency. For operating reports and exception reviews, citation tracing means checking each retrieved quoted passage against the handwritten slip, receipt, or contract to confirm it refers to the same party, the same object, the same period, and the same field. Fields the material does not state remain unknown; authorization or effectiveness clauses are not supplied by assumption.
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, meaning the output path still runs when nothing is retrieved. Retrieval does not perform authorization checks, and the minimum relevance value is a filtering threshold, not a rule that excludes unrelated businesses or unauthorized records. Hybrid retrieval merges multiple result paths through RRF, reranking requires a rerank model enabled in settings, and image retrieval requires corresponding image input and model capability. Full-text retrieval suits exact queries such as numbers and model designations; semantic retrieval suits natural-language questions. An empty citation array does not mean the workflow errored out and stopped; it must be checked together with the downstream handling design.
Hypothetical drill sample: the material text reads, "On 12 March 2024, Stall A issued a handwritten sales order to Buyer B, payment receipt number S-0312, amount four thousand yuan, status received in full." and "On 12 March 2024, Stall A and Buyer B signed a stall lease contract, with the lease term starting 1 April 2024." The draft sentence under review is, "Buyer B completed signing and the stall lease contract took effect on 12 March 2024." That sentence merges the fact of signing with the contract's effective status. The material records only the signing date and states no effective condition or effective date, so the effectiveness portion remains to be verified.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this trial | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Only the stall document base sharing one vector model | Check the vector model identifier first, exclude different models, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Use full-text retrieval for document-number questions and semantic retrieval for natural-language questions, then compare hit differences |
| Minimum relevance | 0.5 | A hypothetical starting point for this trial; compare recall and false recall against 0.3 and 0.7 |
| Citation limit | 3000 tokens | A hypothetical starting point for this trial; compare citation completeness against 1500 and 6000 |
| Reranking | Enabled | Enable when a rerank model is available; compare ordering before and after reranking |
| Question optimization | On | Enable for multi-turn follow-ups or short questions; compare retrieval results before and after completion |
Three common errors
- Treating the signing date as the contract's effective date. The draft sentence says "signed and took effect," while the material records only the 12 March 2024 signing and states no effective condition, so the effective status remains to be verified. The check takes the material text and the draft sentence as input, compares the "signed" and "effective" fields, and marks the effectiveness portion as unsupported when the material states no effective information.
- Treating a retrieval hit as authorization confirmation. Retrieval returns citation content only and does not judge operating permission. The check takes one unauthorized query and one authorized query as input, compares the returned citation arrays, and notes that both may return content, which cannot be used to conclude authorization.
- Treating an empty citation array as an error. The node output length may be zero and the path still runs. The check uses a paper input of an empty array to inspect downstream handling and observes whether the downstream treats it as an empty result, without claiming that retrieval under a fictional rule must return an empty array.
How to verify the setup
- Use the material text and draft sentence from the hypothetical sample as input, compare the "signed" and "effective" fields in the draft sentence against what the material states, and when the material records only the signing date, judge the effectiveness portion unsupported and record the evidence gap.
- Use two knowledge bases sharing one vector model for a linking test, compare citation output before and after linking, confirm that only same-model bases can be linked together, and check identifiers first to exclude different-model bases without fabricating missed-recall or redundant results.
- Use a paper input of an empty array to inspect downstream handling, observe whether the output path runs and whether downstream treats it as an empty result, and make no claim that actual retrieval must return an empty array or that the workflow errored out and stopped.