What the documents look like
Central chain procurement is run by a headquarters supply chain center. The center buys centrally, the distribution center releases goods to stores in waves, and supplier delivery data, WMS distribution notes, store requisition forms and centralized procurement contracts live in separate systems. Contract-to-transaction material comparison checks whether subject, object, quantity, period and authorization status agree across these materials. This page covers only that comparison task and only the knowledge base retrieval and recall step behind it.
The material types differ sharply in structure. Contracts, replenishment plans, distribution notes and store requisition forms carry many precise fields such as identifiers, product names and batch numbers, and relatively little natural language description. That distribution shapes which search mode and parameters are worth testing.
Update frequency is daily or weekly. Retrieval uses materials by their actual ingestion time. No batch synchronization, full synchronization, real-time refresh or transaction-triggered mechanism follows from that frequency.
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 with different vector models cannot be associated together. Before retrieval, check the vector model identifier, exclude non-matching bases and record the observation. Do not run cross-model retrieval to compare missed recall.
Semantic search returns similarity from vector distance. Full-text search depends on keyword hits. Hybrid search merges the two paths through RRF. RRF is a fusion process inside hybrid search, not a switch that can be turned on. The minimum relevance value takes effect only under semantic search or when result reranking is used, and it is a filter threshold. It does not perform permission or entity isolation.
Cited content is output as an array whose length can be 0. When nothing is found, the output path still completes. Image retrieval requires image input and corresponding model capability. Reranking requires a rerank model and must be enabled by configuration. The paths drawn in the overall retrieval diagram do not all execute on every request.
Hypothetical paper exercise sample: a supply chain center holds an excerpt of a centralized procurement contract reading "Party A: Hualing Supply Chain Center; Party B: Ruihe Food Co., Ltd.; subject matter: frozen chicken mid-wings; specification: 1 kg per bag; agreed supply period: 1 March 2024 to 31 March 2024; signature block shows no seals from either party." The same batch contains a distribution note reading "Receiving party: Hualing Distribution Center; product name: frozen chicken mid-wings; quantity: 200 bags; date: 12 March 2024." One draft sentence to be judged reads "Ruihe Food Co., Ltd. completed all supply obligations under this contract in March 2024." That sentence involves two facts, signing and completion of supply, and the source text states neither, so it can only be recorded as unsupported.
Configuration and implementation checks
The table below concerns the sample above and lists only documented settings or implementation checks relevant to this function and task. The values are starting points for this hypothetical exercise, not official defaults.
| Configuration item | Starting point for this hypothetical exercise and reason | Comparison method |
|---|---|---|
| Associated knowledge bases | Include only bases holding the contract and the distribution note that share one vector model | Check whether the vector model identifiers match; exclude those that do not |
| Search mode | Try semantic search and full-text search separately first | Use the sample terms "frozen chicken mid-wings" and "March 2024" and compare what each path returns |
| Citation cap | Start from the combined length of the two sample materials | Test with materials of differing length to see whether cited content stays complete, then adjust |
| Minimum relevance | Mark relevant and irrelevant passages in the sample before setting a threshold | Observe whether relevant passages survive as the threshold changes |
| Question optimization | Enable for multi-turn follow-ups that contain references | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when a rerank model is available | Compare candidate order before and after reranking and record the score range |
Three common errors
- Treating a retrieval hit as authorization or completed signing. The sample contract has no seals in its signature block. Retrieving that contract shows only that the material was ingested; it does not show that the contract took effect.
- Treating minimum relevance as a permission or entity filter. That value is a filter threshold. It does not exclude unrelated companies or unauthorized records, and retrieval by itself does not constitute authorization.
- Forcing knowledge bases with different vector models into one association. The documentation requires associated knowledge bases to share a vector model. When identifiers differ, exclude the base and record the observation. Do not infer any runtime result from a cross-model association.
How to verify the setup
- Input the sample contract text together with the draft sentence "Ruihe Food Co., Ltd. completed all supply obligations under this contract in March 2024." Compare whether the two facts stated in the draft, signing and completion of supply, appear in the source text. The source text records only subject matter, specification and supply period, and states no seals and no completion status. The condition is met when the draft sentence is marked unsupported and the evidence gap is preserved.
- Input the distribution note text with the search term "frozen chicken mid-wings." Compare whether the recalled content contains that product name and the date field. The condition is met when recalled content matches the source text. If recall is empty, handle it under the documented rule that the output array length can be 0, and check the downstream design. Do not claim that a rule-based search necessarily returns an empty array.
- Input two association requests for knowledge bases whose vector model identifiers differ. Compare whether the identifiers match. The condition is met when the non-matching base is excluded and the observation is recorded. Do not continue with cross-model retrieval, and do not assert missed recall or redundant results from it.
This page addresses only central chain procurement, contract-to-transaction material comparison, and knowledge base retrieval and recall. Whether materials, operations or requests are authorized at handover cannot be settled by retrieval or by a form.