What the documents look like
B2B marketplace matching relies on online certification and electronic contracts as its main credentials, with transaction parties and performance records held on the platform side. The document set that copy drafting draws on includes enterprise certification materials, inquiry and quotation sheets, electronic contracts, and platform settlement statements. These materials update daily and per order, so a drafted sentence about price terms, delivery commitments, or settlement conditions needs to be traceable to a specific source passage rather than to a general impression of the account.
The retrieval step carries the job of matching a proposed sentence against the original wording of those materials. A draft may state a certification status, a quotation validity period, or a signing state; the retrieval output is the evidence a reviewer compares against. Retrieval returns relevant content, and it does not perform entity verification, permission isolation, or authorization decisions, which belong to downstream business implementation.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Its output is an array of referenced content, and the length can be zero, meaning the output path still completes when no content is found. Question optimization is optional and follows the configured setting. When enabled, it performs coreference resolution and question expansion before retrieval. Candidates are then recalled through semantic retrieval, full-text retrieval, or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. Result reranking uses a ReRank model to produce a second ordering with a 0-1 score. The minimum relevance setting takes effect only under semantic retrieval or when result reranking is used, and it is a filtering threshold. The reference cap controls, in tokens, the maximum amount of content cited per search.
Hypothetical paper exercise: the source text reads: "Supplier A submitted enterprise certification materials on March 1, 2026, and the certification status is under review; on March 5 of the same month the platform generated quotation sheet XJ-2026-0305, with a quotation validity period ending March 20, 2026; the electronic contract has not been signed." The proposed sentence reads: "This supplier has passed platform certification, and the quotation is valid long term." The point to check is whether the certification status and quotation validity period have been altered by the proposed sentence.
Configuration and implementation checks
| Configuration item | Starting point for this paper exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the knowledge base holding enterprise certification materials, inquiry and quotation sheets, electronic contracts, and platform settlement statements | First check the vector model identifier of each base and exclude bases with a different vector model |
| Search mode | Start with semantic retrieval | Compare recalled content for the same proposed sentence against full-text retrieval and hybrid retrieval |
| Reference cap | Measured in tokens; start at a value that can hold one complete quotation sheet entry | Compare with a smaller value and observe whether the quotation validity field is truncated |
| Minimum relevance | Start at a mid value in the 0-1 range | Label relevant and irrelevant samples first, then compare against higher and lower values |
| Question optimization | Enable for draft requests containing coreference | Compare the original question with the optimized question and check whether the subject was changed incorrectly |
| Result reranking | Enable ReRank when candidates are numerous | Compare ordering before and after reranking and observe whether the material matching the proposed sentence moves forward |
These values are starting points for this paper exercise, not official defaults. RRF is a result fusion process in hybrid retrieval and is not an independent switch.
Three common errors
- Treating a retrieval hit as proof that the entity has been certified. In the sample, the certification status is under review. Retrieving certification materials shows only that the materials exist, and does not support a conclusion that certification passed.
- Treating the quotation validity period as long term. In the sample, the period ends March 20, 2026. The proposed sentence states long-term validity, which conflicts with the source range.
- Writing an unsigned electronic contract as signed. The sample states the electronic contract has not been signed. The proposed sentence does not mention signing status, and that status cannot be assumed.
How to verify the setup
- Input the sample source text and the proposed sentence "This supplier has passed platform certification," then compare the certification status field in the retrieved referenced content. The condition is met only when the referenced content shows certification passed; a status of under review does not meet it.
- Input the proposed sentence "The quotation is valid long term," then compare the validity period field in the referenced content. The condition is met only when the referenced content contains a long-term or open-ended statement; containing March 20, 2026 does not meet it.
- Input the proposed sentence "The electronic contract has been signed," then compare the signing status in the referenced content. The condition is met only when the referenced content shows signed; showing not yet signed does not meet it. If retrieval returns an empty array, check the downstream design against the empty-reference path, and do not infer that the retrieval rule necessarily returns empty.