What the documents look like
Central chain procurement revolves around headquarters purchasing and distribution flow. The documents involved are typically centralized procurement contracts, replenishment plans, distribution orders and store requisition forms. Data comes from ERP purchasing, WMS, store POS and supplier delivery records, and it updates daily or weekly. Errors in this flow lead to store stockouts, inventory buildup or supplier claims.
Product information and selection questions in this category usually ask about the status of a specific item in a specific contract or requisition form. The retrieval step therefore has to carry the subject, the object, the period and the relevant field together into the cited content. When a question asks whether a requisition has been approved, the retrieval result must reflect the actual recorded status, not a plausible assumption about what happens next.
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 using different vector models cannot be associated together; the model identifier must be checked and non-matching bases excluded before any search runs. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch that can be toggled on. The minimum relevance value is a 0–1 filter threshold that takes effect only in semantic retrieval or when result reranking is used. It does not perform permission or entity isolation. The citation limit is measured in tokens and controls how much content a single search may cite. Question optimization performs coreference resolution and question expansion by calling a model before retrieval. Image retrieval requires image input and a model with the corresponding capability, and reranking requires a configured rerank model enabled per settings. These optional paths do not all execute on every request. An empty result is returned as an array whose length may be 0, and the output path still completes.
Hypothetical exercise sample: a chain supply chain center is checking an original centralized procurement contract whose text reads, "Party A and Party B sign a centralized procurement contract for a certain model of refrigerated cabinet; the contract period is January 1, 2025 to December 31, 2025; the agreed supply region is the East China distribution center; settlement is monthly with 60-day terms." A separate original store requisition form reads, "A certain store submitted a requisition on March 5, 2025, for a certain model of refrigerated cabinet, quantity 20 units, status pending review." The draft sentence to be judged is, "The store requisition has been approved and the refrigerated cabinet will be delivered in March 2025." The materials needed for judgment are the contract text and the requisition text above. The authorization boundary is to check the recorded text only, without executing any approval action.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select one base whose vector model identifier matches | Check the model identifier, exclude non-matching bases, and do not run cross-model retrieval |
| Search mode | Test semantic retrieval and full-text retrieval separately first | Compare hit differences between identifier-type questions and natural-language questions |
| Citation limit | Set to a token amount that covers the contract subject and period fields | Test content completeness with actual materials of different lengths, then adjust |
| Minimum relevance | Label relevant and irrelevant samples before setting a threshold | Compare whether cited content retains key fields before and after filtering |
| Question optimization | Enable for follow-up questions containing references; disable for standalone clear questions | Compare whether the subject in the original question is altered incorrectly after optimization |
| Result reranking | Enable per settings when a rerank model is available | Compare rerank scores and vector scores for candidate ordering |
Three common errors
- Treating minimum relevance as a permission or entity isolation mechanism. Input is a query about a store requisition; the comparison item is the threshold setting against returned content; the observable result is that the threshold affects relevance filtering only and does not exclude unrelated business records.
- Describing RRF as a switch that can be turned on. Input is a hybrid retrieval configuration; the comparison item is the documented description against the configuration interface; the observable result is that RRF is a fusion process with no corresponding switch.
- Treating a pending-review status as approved. Input is the requisition text above and the draft sentence; the comparison item is "pending review" against "has been approved"; the observable result is that these are different values of the same field and the draft sentence is unsupported.
How to verify the setup
- Input the contract text and requisition text above; compare "has been approved" in the draft sentence against the "pending review" field in the requisition; the observable result is that the draft sentence is unsupported and the evidence gap must be recorded.
- Input a follow-up question containing a reference; compare the original question against the retrieval terms after question optimization; the observable result meets the condition only when the subject is not altered incorrectly, and requires correction when the subject is replaced.
- Input a clearly labelled paper input of an empty result array []; compare how the downstream design handles the empty array; the observable result is that the output path completes with a citation length of 0, without claiming that a fictional rule search necessarily returns [].