What the documents look like
B2B marketplace matching transactions rely on four document families for operating reports and exception reviews: enterprise verification materials, inquiry and quotation sheets, electronic contracts, and platform settlement statements. Online verification and electronic contracts appear more often in this category than in wholesale market stall trading, and transactions do not depend on a physical stall. The trading entity, contract status, and settlement amount are spread across separate documents. Retrieval for review work requires comparing the same entity, the same period, and the same field across those documents.
Updates arrive daily and per order. A retrieval result reflects the index state established when the material was ingested. If a verification record was loaded before a status change, the retrieved content shows the earlier state. Retrieval does not refresh source systems, and it does not confirm that a document is current.
The primary task on this page is retrieval for operating reports and exception reviews. The review compares retrieved source text against a draft sentence or a handover request. Retrieval supplies candidate content. It does not verify enterprise identity, approve a contract, or authorize a payment.
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. The vector model identifier is checked first, and non-matching bases are excluded. No cross-model retrieval is performed to compare recall differences.
Semantic retrieval depends on vector distance. Full-text retrieval depends on keyword hits. Hybrid retrieval combines both paths through RRF. Result reranking applies a ReRank model for a second ordering. The citation limit controls how many tokens of content a single search may cite. The minimum relevance value applies only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval.
Retrieval does not perform permission or entity isolation. A similarity score is not a filter for enterprise identity, authorization status, or contract validity. An empty result is returned as an array whose length may be zero, and the downstream path still completes.
Hypothetical paper exercise sample. Source text: "Supplier A submitted enterprise verification materials in March 2025, and the review status is pending review. The inquiry and quotation sheet for the same period states that the quotation is valid until April 30, 2025. The electronic contract text names Supplier A and Buyer B as signing parties, and the signing date field is empty." Draft sentence to be assessed: "Supplier A has completed verification, and the contract has taken effect." The comparison checks the review status field in the verification materials against "has completed verification," and checks the signing date field in the contract against "has taken effect."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate one knowledge base that uses a single vector model | Same vector model is required for association; check the vector model identifier first and exclude non-matching bases |
| Search mode | Run semantic retrieval and full-text retrieval once each | Compare recall differences between natural-language questions and identifier-type questions |
| Citation limit | Set to a token amount covering the full text of the longest single settlement statement | Compare against the shortest inquiry and quotation sheet to observe whether content is truncated |
| Minimum relevance | Leave disabled and observe raw recall first | Decide later whether to use it for filtering after relevant and irrelevant samples are labeled |
| Question optimization | Leave disabled and test with complete questions | Compare against the completed result after enabling, and check whether the entity was altered |
| Result reranking | Leave disabled for now | Evaluate after candidate volume and question completeness are confirmed |
Three common errors
- Treating a pending review status as completed verification. The input is "review status is pending review" in the verification materials. The comparison object is the draft sentence "has completed verification." Both refer to the same entity and the same field, and the values are mutually exclusive, which makes this a contradiction. The appearance of that material in retrieval results does not mean verification passed.
- Treating named signing parties as contract effectiveness. The input is the electronic contract with an empty signing date field. The comparison object is the draft sentence "the contract has taken effect." Signing, reaching an agreement, and contract effectiveness are three distinct facts. When effectiveness evidence is absent, the item is recorded as pending verification.
- Treating quotation validity as a supply deadline. The input is the inquiry and quotation sheet stating "quotation valid until April 30, 2025." The comparison object is any statement about a supply cutoff date. These are different fields, and quotation validity does not indicate a supply deadline.
How to verify the setup
- Input the verification materials from the hypothetical sample and the draft sentence "Supplier A has completed verification." Compare the review status field value against the draft wording. When retrieval hits that material and the review status shows pending review, the draft sentence is not supported. If retrieval does not hit the material, record an evidence gap and do not treat it as a denial.
- Input the electronic contract from the hypothetical sample and the draft sentence "the contract has taken effect." Compare the signing date field and any effectiveness-related field. When the signing date field is empty and no other effectiveness record exists, the draft sentence is pending verification. Do not supply authorization or effectiveness clauses to obtain a definite answer.
- Input a clearly labeled paper empty result
[]and check whether the downstream design carries the path through with a citation length of zero. This check only verifies how an empty result is handled. It does not claim that any invented rule always returns[].
These checks use the same entity, the same period, and the same field. A contradiction requires mutually exclusive values in the same field. An unstated value, or evidence that shows only a related stage, is recorded as unsupported or pending verification. Pending review does not demonstrate the absence of an operating permission. Materials needed for a check must sit inside an accessible target retrieval scope or a clearly defined manual input. When they are missing, record the gap.