What the documents look like
Supermarket supplier onboarding and settlement revolves around four document families: procurement contracts, delivery acceptance notes, rebate agreements, and reconciliation statements. The data arrives from ERP procurement, supplier portals, POS sales, and finance systems, with weekly or monthly update cycles. Compared with large-format concession counters, supplier onboarding settlement is mostly purchase-sale or consignment payables, and does not necessarily involve on-site operations. For cross-border materials and multilingual retrieval, this means a citation must be able to return to a specific clause, quantity, or amount field in one of those four document types.
The citation tracing step in this category answers a narrow question: can a multilingual retrieval result be mapped back to a specific clause, quantity, or monetary field in a named document, and which version of that document does the citation belong to. Retrieval does not constitute authorization, and it does not replace financial reconciliation. A similarity score is a relevance number; it is not a category filter, an entity filter, or a permission filter.
Two features of this category shape the tracing requirement. First, cross-border materials arrive in more than one language, so a Chinese question may need to match an English contract clause. Second, the same commercial fact can appear in several documents at different stages, so tracing must distinguish the contract text from the reconciliation text rather than treating them as one source.
Constraints at the citation tracing step
The knowledge base search node outputs cited content as an array, and the array length can be zero. When the array is empty, the output path still completes. Knowledge bases can only be associated together when they share the same vector model. Semantic retrieval supports cross-language understanding, for example a Chinese question matching an English knowledge point, which is useful for cross-border materials. The similarity value is a relevance number and does not perform permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF, and reranking can produce a 0-1 score used for filtering. The minimum relevance setting takes effect only under semantic retrieval or when result reranking is used.
A hypothetical paper exercise sample follows. A supplier portal uploads an English procurement contract revision page containing two sentences: "Delivery tolerance for frozen goods is plus or minus 3 percent of ordered quantity." and "Rebate rate for Q3 is 2 percent of net purchases." A Chinese reconciliation statement records "Q3 rebate calculated at 2 percent of net purchases." A draft sentence awaiting review reads: "The supplier's Q3 rebate rate is 2 percent, and accepted quantity allows a plus or minus 3 percent deviation." The subject required for the judgment is the supplier, the objects are the rebate rate and the acceptance tolerance, the effective scope is Q3, and the fields are rebate rate and acceptance tolerance. This sample is a paper exercise and does not involve a real customer or an actual test.
The contract revision page states a Q3 rebate rate and a delivery tolerance for frozen goods. It does not state that the revision has been signed, that cooperation has been reached, or that the contract has taken effect. Signing, reaching cooperation, and contract effectiveness are three distinct facts. The reconciliation statement records the Q3 rebate calculation but does not carry the delivery tolerance. Any field not stated in the supplied text remains unverified.
Configuration and implementation checks
| Configuration or check item | Purpose on this page | Hypothetical starting value and reason |
|---|---|---|
| Associated knowledge bases | Limit the retrieval scope | Select knowledge bases containing the procurement contract and the reconciliation statement; first verify that the vector model identifiers match |
| Search mode | Determine the recall path | Semantic retrieval for cross-language matching; use full-text or hybrid retrieval for numbered fields |
| Citation cap | Control the total cited tokens | Hypothetical starting value set to cover one contract revision page plus one reconciliation passage; adjust after comparing with single-document length |
| Minimum relevance | Filter low-relevance results | Hypothetical starting value at a middle point; compare with the score of the draft sentence hitting the source text and observe whether the target citation is filtered out |
| Result reranking | Secondary ordering | Used when a rerank model is available and enabled by configuration; compare the rerank score with the vector score |
| Cited content output | Hand off downstream | Check whether the citation array is passed to the answering step; when the array is empty, check the downstream design |
The citation cap is expressed in tokens, not in a guessed item count. The minimum relevance value is a filter threshold, not a final ordering rule. After question optimization, hybrid retrieval, image retrieval, or result reranking is enabled, the final result may be merged from multiple recall paths and does not necessarily follow a single vector similarity order. Image retrieval requires a corresponding image input and model capability. Reranking requires a rerank model and must be enabled by configuration. The optional paths in the overall retrieval flow do not all execute on every request.
Three common errors
- Treating a semantic retrieval relevance score as a permission or entity filter. The observable error is assuming that raising the minimum relevance value excludes other suppliers or unauthorized records. An executable check is to submit the same question and compare whether the returned citations still contain different supplier subjects, then observe whether the score corresponds to the subject field. A relevance score is not a category, entity, or permission filter, and retrieval is not authorization.
- Associating knowledge bases that use different vector models. The observable error is selecting two knowledge bases with different vector model identifiers and expecting normal retrieval. An executable check is to verify each knowledge base's vector model identifier, exclude any that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls; only verify identifiers, exclude non-matching bases, and record what is observed.
- Treating a draft sentence under review as an effective clause. The observable error is a draft sentence stating "the rebate rate is 2 percent" when the source text only records the Q3 rebate rate and does not record signing or effectiveness. An executable check is to compare the draft sentence fields with the effective scope stated in the source text, confirm that signing, reaching cooperation, and contract effectiveness are three distinct facts, and mark unstated parts as pending verification. A pending review does not establish that operation permission is absent.
How to verify the setup
- Input the English contract revision page text and the Chinese reconciliation statement text from the hypothetical sample. Compare whether the retrieval citations cover both the rebate rate field and the acceptance tolerance field, and observe whether each citation corresponds to a specific source sentence. The condition is met when the citations can be traced back to specific source sentences. Do not guarantee a first-hit result, full recall, or that an unrelated query returns empty.
- Input the draft sentence "The supplier's Q3 rebate rate is 2 percent, and accepted quantity allows a plus or minus 3 percent deviation." Compare the draft sentence fields with the effective scope stated in the source text, and observe whether the rebate rate carries a Q3 qualifier and whether the acceptance tolerance carries a subject and goods-category qualifier. The condition is met when the draft sentence fields match the source text and no effectiveness status absent from the source text has been added. A judgment requires the same subject, the same period, and the same field; only mutually exclusive values in the same field constitute a contradiction. A field that is not stated, or that has evidence only for a related stage, is unsupported or pending verification.
- Input an explicit paper empty result array []. Compare whether the downstream answering step still executes, and observe whether the workflow breaks because of the empty array. The condition is met when the downstream design can accept the empty array and record the gap. Do not claim that a hypothetical rule retrieval necessarily returns []. The search node outputs a citation array whose length can be zero, and an empty array does not mean the workflow has errored and terminated.
These checks describe expected observations and do not claim that any test has been performed. Materials needed for verification must be inside an accessible target retrieval scope or an explicit manual input; when they are missing, record the gap.