What the documents look like
Enterprise certification materials, request-for-quotation sheets, electronic contracts and platform settlement statements form the main document sources for order fulfilment and after-sales question handling on a B2B marketplace. These materials typically carry fields such as enterprise name, unified social credit code, quotation validity period, contract signature status and settlement cycle. Compared with stall trading in wholesale markets, B2B platform transactions rely more on online certification and electronic contracts and less on an on-site stall, so retrieval requests often mix natural-language descriptions with identifiers. The review status of enterprise certification materials, the signature and effectiveness status of an electronic contract, and the validity period of a quotation sheet are facts held in different fields. A retrieval result cannot treat the value of one field as the conclusion of another.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. During implementation, check the vector model identifier first, exclude those that do not match, and do not run cross-model retrieval to compare missed recalls. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, and hybrid retrieval merges the two paths through RRF. The minimum relevance score is a filter threshold and takes effect only in semantic retrieval or when result reranking is used; it does not perform permission or entity isolation. The reference limit is measured in tokens and controls the maximum amount of content referenced per search. Question optimisation adds one model call before retrieval for coreference resolution and question expansion. The output reference content is returned as an array and its length can be 0; when nothing is retrieved, the output path still completes. Image retrieval requires the corresponding image input and model capability, and reranking requires a model and must be enabled by configuration. The optional paths shown in the document flow diagram do not mean every request executes all of them.
Configuration and implementation checks
Hypothetical drill sample: a B2B platform operator handles an after-sales enquiry for one matched order. The source text available for comparison reads: "Supplier A submitted enterprise certification materials on 10 March 2025, certification status pending review; quotation sheet number RFQ-20250310-07, quotation valid until 31 March 2025; electronic contract number CT-20250312-11, both signature fields blank; platform settlement statement number ST-20250315-03, settlement status unsettled." The draft sentence to be assessed reads: "This order has completed enterprise certification, the contract has taken effect, and payment can be initiated against the settlement statement."
| Configuration item | Value in this sample | Reason and comparison method |
|---|---|---|
| Associated knowledge base | Order fulfilment and after-sales knowledge base | Check the vector model identifier first and confirm it matches the current node before associating |
| Search mode | Hybrid retrieval | The sample contains identifiers and natural language; compare the semantic and full-text result paths |
| Reference limit | 3000 tokens as the hypothetical trial starting point | Compare against the combined length of the four sample materials and observe whether references are complete |
| Minimum relevance score | 0.3 as the hypothetical trial starting point | Label relevant and irrelevant samples first, then compare against filtered results |
| Question optimisation | Enabled | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enabled by configuration | After confirming a reranking model is available, compare rerank scores with vector scores |
These values are hypothetical trial starting points for this drill and are not official default values.
Three common errors
- Treating "pending review" as "certified". In the sample, the status field of the enterprise certification materials is pending review, while the draft sentence states that enterprise certification has been completed. The two values are mutually exclusive within the same field, and the sentence is not supported.
- Treating "signature fields blank" as "contract effective". In the sample, both signature fields of the electronic contract are blank. Signing, reaching agreement and contract effectiveness are three distinct facts, and the draft sentence's claim that the contract has taken effect lacks a basis.
- Treating "quotation validity period" as "supply period". In the sample, the quotation is valid until 31 March 2025. That field only describes the range within which the quotation is valid and does not indicate a supply period; initiating payment on that basis is a cross-field inference.
How to verify the setup
- Input the sample source text and the draft sentence, compare the certification status in the draft sentence with the certification status field in the source text, and observe whether the result identifies that "pending review" and "completed enterprise certification" are mutually exclusive within the same field. When this condition is met, the comparison conclusion should mark the sentence as unsupported.
- Input the sample source text and the draft sentence, compare the contract status in the draft sentence with the signature field in the source text, and observe whether the result identifies that blank signature fields do not constitute contract effectiveness. When this condition is met, the comparison conclusion should retain the evidence gap and not supply authorisation or effectiveness clauses.
- Input the sample source text and the draft sentence, compare the payment basis in the draft sentence with the quotation validity period and settlement status fields in the source text, and observe whether the result identifies that a quotation validity period does not indicate a supply period and that an unsettled status does not indicate that payment may proceed. When this condition is met, the comparison conclusion should record the items still to be verified.
When the paper input is an empty array [], check whether the downstream design accepts a reference output of length 0; do not claim that a fictitious rule retrieval necessarily returns []. Retrieval or a form does not constitute authorised execution, a pending review does not demonstrate a lack of operating permission, and an unsigned termination does not constitute a completed termination.