What the documents look like
Wholesale market stall trading runs on on-the-spot bargaining and consignment shipping. The paperwork is fragmented: purchase slips, handwritten sales notes, stall lease contracts, and payment receipts sit with different parties, and the same consignment can carry different product names, quantities, and amounts across those documents. For cross-border material and multilingual retrieval, citation tracing requires that an answer can be taken back to the exact wording of a specific document. A model summary is not enough for reconciliation.
Documents change daily and per transaction. When the same stall issues several versions of a slip in a short window, a check has to confirm that the cited version matches the version currently in effect. Retrieval returns references; it does not decide which version governs.
Constraints at the citation tracing step
The knowledge base search node outputs references as an array, and the length can be zero. When the array is empty, the output path still completes. A search that finds nothing is therefore not a broken workflow, and the downstream design has to handle an empty reference set on its own.
Knowledge bases selected together must use the same vector model; bases with different vector models cannot be linked together. Semantic retrieval supports cross-language matching, and a similarity score is a relevance measure. It does not act as a permission filter, an entity separator, or a business ownership test. Hybrid retrieval merges multiple result paths through RRF, and reranking produces a 0–1 score used for filtering. These belong to ranking and filtering, and they say nothing about whether a document was authorized for use.
Hypothetical drill sample. A stall owner and a buyer disagree about the balance on a consignment. The source material available for checking consists of two sentences. First: "The handwritten sales note dated 12 March 2025 states: Style A fabric, 200 metres, unit price 18 yuan, total 3,600 yuan, 2,000 yuan already received, remaining 1,600 yuan to be settled within three days of signing the consignment note." Second: "The logistics consignment note dated 14 March 2025 states: signed for by Li, date of signing 14 March 2025."
The draft sentence to be judged is: "The buyer signed for the goods on 14 March 2025, and the remaining 1,600 yuan should be settled by 17 March 2025." This draft depends on three fields matching at once: who signed, the signing date, and the starting point of the balance deadline. The source material supports the signing date and the identity of the signer. It does not state how the three-day period is counted, so the 17 March date is not supported by the quoted text.
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Relation to the sample |
|---|---|---|
| Linked knowledge bases | Check the selected bases and the document versions; only bases with the same vector model can be linked | Confirm whether the sales note and the consignment note are searchable together under one vector model |
| Search mode | Semantic, full-text, or hybrid retrieval chosen to match the material | The handwritten sales note carries amounts and dates, so full-text retrieval can help with exact matching |
| Reference limit | Each search cites at most n tokens of content | Hypothetical trial starting value 800 tokens, chosen because a single slip is short; compare against the combined length of both documents before adjusting |
| Minimum relevance | A 0–1 value, effective only under semantic retrieval or reranking | Hypothetical trial starting value 0.3, used to observe whether low-relevance fragments are filtered; compare against the content actually recalled |
| Question optimisation | Coreference removal and question expansion before retrieval | If the user asks "what about that balance", confirm whether the optimised question keeps the original document subject |
| Reference output | Array format, length may be zero | The downstream handling of an empty array has to be checked separately and cannot be treated as a successful search |
Three common errors
- Treating retrieval output as proof of authorization or settlement. A returned reference only means relevant text was matched. It does not mean the balance was settled or that a signing was authorized. The check compares the subject, date, and amount fields in the quoted source text; fluent wording does not replace source evidence.
- Treating the similarity score as an entity separator. The minimum relevance value is a filtering threshold and does not distinguish stalls, buyers, or document types. The check confirms whether the cited content comes from a valid document for the target stall, and records an evidence gap when access rights are missing.
- Treating an empty reference array as a workflow failure. The search node output length may be zero while the output path still completes. The check uses an explicit paper input of [] to inspect the downstream design and confirm how an empty result is signalled or escalated. It does not claim that a fictional rule search necessarily returns [].
How to verify the setup
- Enter the sales note text and the consignment note text from the hypothetical drill sample above. Compare the draft sentence "The buyer signed for the goods on 14 March 2025, and the remaining 1,600 yuan should be settled by 17 March 2025" field by field against the source. The signing date and signer are supported. The starting point of the balance deadline is not stated in the quoted text, so the 17 March date remains unsupported and the gap is recorded.
- Enter two knowledge base identifiers that use different vector models and compare whether the system permits linking them. If the identifiers do not match, exclude the non-conforming base and record the observation. Do not run a cross-model search to compare missed recalls, and do not infer what would happen after linking.
- Enter a paper input of an empty reference array [] and compare whether the downstream node handles the empty result as designed. If the downstream node treats the empty array as a fatal error, the handling logic needs revision. If the downstream path completes and emits a notice, the condition is met. This check does not claim that a fictional rule search necessarily returns [].
At operational handover, a pending review status does not show that operating permission is absent. Signing, reaching an agreement, and a contract taking effect are three distinct facts. A quotation validity period does not indicate a supply period. Material needed for a reconciliation must sit inside an authorized target retrieval scope or an explicit manual input; when it is missing, record the gap.