What the documents look like
B2B marketplace matching runs on online company certification materials, inquiry and quotation sheets, electronic contracts, and platform settlement statements. Company identity and fulfillment records sit across the platform itself, the national enterprise credit information publicity system, payment records, and logistics traces. In cross-border and multilingual retrieval, the same supplier may submit certification material in Chinese and in English, and the model numbers, trade terms, and settlement currencies on a quotation sheet become precise retrieval targets. These materials update daily or per order. The retrieval step has to let a question asked in one language locate the corresponding source text in another.
Three document families anchor later reconciliation: company certification materials, electronic contracts, and settlement statements. Their party names, validity periods, and amount fields are the fields a reviewer compares against. The relative frequency of online certification and electronic contracts in this category means fewer transactions depend on an on-site stall, so the retrieved document text, not a physical visit, carries the evidence.
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. Check the vector model identifier first, exclude any base that does not match, and record what was actually observed. Do not infer the runtime result of associating mismatched bases.
Semantic retrieval supports cross-language understanding: a Chinese question can match an English knowledge point, which suits multilingual retrieval over quotation sheets and certification materials. Full-text retrieval depends on keyword hits and suits precise queries such as model numbers and contract numbers. Hybrid retrieval uses semantic recall and full-text recall together and merges them through RRF; RRF is a fusion process, not a switch that can be turned on. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold. It does not perform permission or entity isolation. Retrieval outputs references as an array whose length can be 0, and the output path still runs when nothing is found. Image retrieval requires an image in the input and the corresponding model capability, and reranking requires a configured model enabled by its setting; these paths do not run on every request.
Configuration and implementation checks
Hypothetical drill sample: a B2B platform operator handles one supplier certification record. The source text reads "Supplier A Co., Ltd., unified social credit code example value, certification status pending review, certification material submission date 10 March 2025, certification valid until 9 March 2026." The draft sentence to be judged reads "Supplier A Co., Ltd. has completed platform certification, with a validity period from 10 March 2025 to 9 March 2026." The configuration below is set around this sample's retrieval and reconciliation needs.
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge base | Associate one knowledge base holding certification materials | Check that base's vector model identifier against the node requirement before associating |
| Search mode | Try hybrid retrieval first | Compare the recalled content for the same question against semantic retrieval and full-text retrieval separately |
| Reference limit | Start at a token amount covering the complete fields of one certification record | Test with certification records of different lengths to see whether fields are truncated, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set a threshold | Takes effect only under semantic retrieval or reranking; compare results before and after filtering |
| Question optimization | Enable for continuous follow-up questions containing references | Compare the original question with the optimized question and check whether the subject was changed incorrectly |
| Result reranking | Enable when candidates are numerous | Requires a configured rerank model; compare the rerank score with the vector score |
Three common errors
- Treating a pending-review status as completed certification. The sample source says "certification status pending review," while the draft sentence says "has completed platform certification." The same subject and the same field carry mutually exclusive values, so the draft sentence is unsupported. To check, input the source text and the draft sentence, compare the certification status field, and treat a mismatch between the draft and the source as failing the condition.
- Treating retrieval recall as authorization to execute. Recalling a certification record only means the content was retrieved; it does not mean certification passed or a contract took effect. To check, input the retrieval output and the business request, compare the authorization state the request requires against the retrieval result, and treat a retrieval result without an authorization conclusion as unable to advance the business action.
- Forcing an association across vector models. The node requires associating knowledge bases with the same vector model. When a base with a different vector model appears, check its identifier and exclude it rather than running a cross-model search to compare missed recall. To check, input each base's vector model identifier, compare whether they match, and leave mismatched bases out of the association.
How to verify the setup
- Input the hypothetical source text and the draft sentence, then compare the certification status field. The source reads "pending review" and the draft reads "has completed platform certification." These are mutually exclusive, so the draft sentence is judged unsupported. If the draft instead reads "certification status pending review," matching the source, the condition is met.
- Input a follow-up question containing a reference, then compare the subject before and after question optimization. If the optimized question replaces Supplier A with another subject, the optimization result is judged wrong. If the subject stays consistent, the condition is met.
- Input a paper empty result
[]and check whether the downstream design accepts an empty reference. The node's reference output length can be 0 and the output path still runs. Do not claim that a fictional rule search necessarily returns[]; record only the output actually observed.
At operational handover, certification material sits in pending-review status. Retrieval or form collection does not constitute certification approval, contract effectiveness, or authorized execution. Signing, reaching cooperation, and contract effectiveness are three distinct facts, and a quotation validity period does not indicate a supply period. Material needed for reconciliation must sit within an accessible target retrieval scope or an explicit manual input; when it is missing, record the gap.