What the documents look like
B2B marketplace matching transactions depend on online certification and electronic contracts, and they do not rely on a physical stall. The working materials for this category include enterprise certification files, inquiry and quotation sheets, electronic contracts, and platform settlement statements. These materials update daily or per order. During operations handover, the current state of each material and the operations that the next shift may perform on it must be stated explicitly. Knowledge base retrieval and recall locate materials and return quoted content. The retrieval result does not grant permission, does not record approval, and does not execute a business action.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The search output is an array of quoted content, and its length can be zero, which means the downstream link still runs even when no content is found. Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. A ReRank model performs a second ordering of candidate text. The minimum relevance value takes effect only under semantic retrieval or when result reordering is used, and it filters results by score rather than setting the final ranking order. The quote limit is measured in tokens. Question optimization performs coreference resolution and question expansion. Image retrieval requires image input and the corresponding model capability. ReRanking requires a ReRank model and must be enabled according to the configuration. Knowledge bases that use different vector models cannot be associated together; the model identifier must be checked first and non-matching bases excluded. Retrieval does not perform permission or entity isolation, and a similarity score does not constitute an authorization judgment. The same-model requirement for associated knowledge bases does not extend to the national enterprise credit information publicity system or to any other external system.
Hypothetical paper exercise sample: the source material reads, "Party A, XX Supply Chain Management Co., Ltd., submitted enterprise certification files on 10 March 2025, and the status is pending review. The inquiry and quotation sheet of Party B, XX Trading Co., Ltd., is numbered XJ-20250310-01, and the quotation is valid until 31 March 2025; the material does not record acceptance or rejection. Electronic contract HT-20250310-01 has not yet been signed by both parties." The draft sentence to be assessed reads, "Party A's certification has passed, Party B's quotation has been accepted, the contract has taken effect, and settlement can proceed." The handover request reads, "Please hand over the retrieval results of the above materials to the next shift of operations staff."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate the enterprise certification base, the inquiry and quotation base, and the electronic contract base that use the same vector model | Check that the vector model identifiers match before associating, then check whether the target quotation material is in the base and whether the current account has access, and exclude bases whose identifiers do not match or that the account cannot access |
| Search mode | Hybrid retrieval | Compare recall differences for the sample question under semantic, full-text, and hybrid modes; numbered questions can be compared against full-text results |
| Quote limit | Hypothetical starting point of 2000 tokens for this exercise | Test content completeness with materials of different lengths, and compare whether quoted content is truncated against smaller or larger values |
| Minimum relevance | Hypothetical starting point of 0.3 for this exercise | Label relevant and irrelevant samples first, then compare whether irrelevant content is filtered at different thresholds |
| Question optimization | Enable for multi-turn follow-up questions that contain references | Compare the original question with the optimized question, and check whether the subject has been changed incorrectly |
| ReRank | Enable when a ReRank model is available | Compare candidate text ordering with and without reordering, and confirm that the model is available |
Three common errors
- Treating retrieval results or the draft sentence as an authorization credential. The input is the retrieval result for the sample materials together with the draft sentence. Compare the certification status, the quotation material record, and the contract signature state. Pending review and passed are distinct states, so a draft sentence that claims certification has passed contradicts the source. The quotation material does not record acceptance or rejection, so a claim that the quotation has been accepted is unsupported and remains to be verified; missing acceptance evidence does not establish rejection and does not establish an explicit contradiction. Retrieval returns quoted content only, and it does not execute approval or signing.
- Associating knowledge bases that use different vector models. The input is the vector model identifier of each base. Compare whether the identifiers are the same. If they differ, the bases cannot be associated together; exclude the non-matching base and record the actual observation. The same-model requirement does not extend to the national enterprise credit information publicity system or to any other external system.
- Treating minimum relevance as a permission or entity filter. The input is the similarity score and the business entity identifier. Compare the score with the relevance threshold. Relevance affects content ordering and filtering only, and it does not exclude unrelated enterprises or unauthorized records. Before retrieval, check whether the target quotation material is in the base and whether the current account has access; if the material or the access is absent, record the gap.
How to verify the setup
- Input the sample source material and the draft sentence. Compare the certification status, the quotation record, and the contract signature state in the draft sentence with the source material. If the draft sentence states that certification has passed, it contradicts the recorded pending-review state. The source records incomplete signing and provides no basis for contract effectiveness, so a claim that the contract has taken effect is unsupported and remains to be verified. If the draft sentence states that the quotation has been accepted while the source records neither acceptance nor rejection, mark the claim as unsupported and pending verification. The expectation is that the draft sentence restates the states recorded in the source only.
- Input the vector model identifiers of the knowledge bases together with the record showing whether the target quotation material is in the base and whether the current account has access. Compare whether the identifiers match, whether the material exists, and whether the account has access. If the identifiers differ, or the material is missing, or the account lacks access, the setting cannot be promised as executable, and the gap is recorded. The expectation is that only bases with matching identifiers and with access are associated.
- Input the sample question together with an empty-result downstream design. Compare whether the length of the retrieval output quoted-content array is zero and whether the downstream link runs. If the output is an empty array and the downstream design can accept it, the behavior matches the documented empty-result behavior. The expectation is that a hit is not guaranteed on every search and that the empty-result link still runs.
The hypothetical starting values above are exercise assumptions only. They are not official defaults, and no test, deployment, approval, refund, or other business action is claimed to have occurred. Fields not provided in the source material remain unknown.