What the documents look like
In central chain procurement, knowledge base retrieval serves supply chain centers, procurement managers and distribution center staff. Retrieval demand concentrates on four material types: centralized procurement contracts, replenishment plans, distribution orders and store requisition orders. These materials share dense document numbering, explicit subject and period fields, and a limited set of status values. ERP procurement, WMS, store POS and supplier delivery data update daily or weekly, and retrieval results support store service replies and internal training answers.
Number and status fields determine the retrieval method. Contract numbers, distribution order numbers and requisition order numbers are exact-query objects; semantically similar records with different numbers do not constitute the same object. The minimum relevance value is a filter threshold that takes effect under semantic retrieval or result reranking; that value does not perform subject isolation or permission judgment. The knowledge base search node can associate one or more knowledge bases that use the same vector model; knowledge bases with different vector models cannot be associated together.
Constraints at the retrieval step
The retrieval step only returns reference content, output as an array, whose length can be 0. Whether a store requisition order has been approved, whether a distribution order has been handed over, and whether a contract has taken effect are all business status facts; a retrieval hit on a related passage does not change those facts. The reference limit is measured in tokens and controls the total volume of reference content in a single search; materials with large length variation require checking content completeness before adjusting it.
Hypothetical paper exercise sample: the source material reads "Distribution order PS-2024-0317, receiving store is East China Store Three, planned arrival date March 20, 2024, current status: dispatched, handover status: not signed for." The proposed determination sentence is "This distribution order has completed store signing." The comparison object is the handover status field; the source value is not signed for, and the proposed sentence claims signing is complete, so the same field holds mutually exclusive values. This sample only illustrates the comparison method and does not represent a real business record.
Configuration and implementation checks
| Configuration item | Documented setting or check item | Relation to this page's sample |
|---|---|---|
| Associated knowledge bases | One or more knowledge bases using the same vector model can be associated | Check the vector model identifier first and exclude inconsistent ones |
| Search mode | Semantic retrieval, full-text retrieval and hybrid retrieval are selectable | For number-type questions, compare full-text retrieval hits first |
| Reference limit | Unit is token; controls total reference content | Hypothetical trial starting point 2000, adjusted after comparing with sample passage length |
| Minimum relevance | Value from 0 to 1; effective only under semantic retrieval or result reranking | Label relevant and irrelevant samples first, then compare threshold effects |
| Question optimization | Used for coreference resolution and question expansion | Compare whether the subject in the original and optimized question stays consistent |
| Result reranking | Requires a rerank model and must be enabled by setting | When not enabled, this path does not participate in this retrieval |
The hypothetical starting point of 2000 tokens for the reference limit was chosen because the sample passage length is close to other order passages in the same knowledge base, which makes truncation positions comparable; this value is a paper trial starting point for this page and is not an official default.
Three common errors
- Treating a retrieval hit as status confirmation. Input is the distribution order text above; retrieval returns that passage. Outputting "signing is complete" on that basis is wrong because it equates a recalled result with the value of the handover status field. The check point is the source value of the handover status field.
- Associating knowledge bases with different vector models together. Input is two knowledge bases with different vector model identifiers; merging retrieval directly ignores the same-vector-model requirement. The check point is whether the vector model identifiers of the knowledge bases match.
- Writing a pending-review status as already effective. Input is a clause passage in a centralized procurement contract marked "pending review"; a draft sentence reading "this clause has taken effect" is wrong because it treats an under-review status as an effective fact. The check point is whether the contract status field and the draft sentence claim share the same field and value.
How to verify the setup
- Input the distribution order text above and the proposed determination sentence "This distribution order has completed store signing." Compare the handover status field value with the proposed sentence claim. If the field is not signed for while the proposed sentence claims signing, the determination is unsupported; only if the field is signed for does it agree with the proposed sentence.
- Input two knowledge base identifiers with different vector model identifiers and compare whether the identifiers match. If they do not match, exclude joint association and record the actual observation; do not run cross-model retrieval to compare recall differences.
- Input an empty reference array [] and check whether the downstream chain continues under the condition that length can be 0. This check examines the output structure and does not claim that a fabricated rule retrieval necessarily returns an empty array.
All three verification items address this page's sample and configuration items and have not been executed; conclusions rest on paper comparison.