What the documents look like
Fresh-food concession counters in hypermarkets are frequently run by on-site supplier staff under a revenue-share arrangement. The working records for such a counter are the concession contract, daily sales reports, loss reports, and food incoming inspection records. These records carry permission facts as well as numbers: who may sign off on a loss confirmation, which batch of incoming inspection evidence belongs to which supplier, and which handover actions remain pending review. The retrieval step only returns text from material already stored in a knowledge base. It does not replace an authorization decision, and a retrieved sentence is not an approval.
The material updates daily or per batch. A retrieval result reflects the content as stored at indexing time. The search node does not carry a real-time synchronization duty, and this page does not assign one to it.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. The identifier of each base's vector model is checked first, and any base that does not match is excluded. A cross-model search is not executed for comparison.
The search mode can be semantic retrieval, full-text retrieval, or hybrid retrieval. Hybrid retrieval combines multiple result paths through RRF. RRF is a fusion step in that combination, and it is not a switch that is turned on. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold. It does not act as a permission filter or an entity isolation rule, and it does not exclude unrelated businesses or unauthorized records. The reference cap is measured in tokens. Question optimization performs coreference resolution and question expansion, and it adds an extra model call before retrieval. The referenced content is returned as an array whose length can be zero, so the output path still runs when nothing is retrieved.
Image retrieval requires an image input and the corresponding model capability. Reranking requires a rerank model and must be enabled according to the settings in use.
Hypothetical paper exercise sample: a store fresh-food department and concession supplier A sign a concession contract stating that loss is shared between the two parties in proportion. The paper material reads: "Supplier A's on-site staff complete the initial entry of the daily loss report; the store fresh-food department supervisor reviews and then confirms it; a loss report that has not been reviewed may not enter settlement." A separate handover request draft reads: "Please confirm supplier A's loss report from yesterday directly and include it in the revenue-share settlement." The draft carries no review record and does not name a reviewer.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the bases holding the concession contract and loss reports that share one vector model | Check each base's vector model identifier first, exclude non-matching bases, and do not run a cross-model search |
| Search mode | Try semantic retrieval and full-text retrieval separately on a question about loss report review | Compare the content recalled by both modes for the same question and observe which type of material is hit |
| Reference cap | Set a hypothetical starting value large enough to hold one loss report clause in tokens | Test with material of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Label related and unrelated samples first, then set one value to be compared | It takes effect only under semantic retrieval or result reranking; compare candidates before and after filtering |
| Question optimization | Enable for short multi-turn follow-up questions, disable for questions that are already explicit | Compare the original question with the optimized question and check whether the subject was altered |
| Reranking | Enable according to whether a rerank model is actually available | After enabling, compare the rerank score with the vector score and observe the change in ordering |
Three common errors
- Treating a retrieval hit as a completed authorization. The input is the retrieval result for the question about loss report review. The comparison is whether that result contains a reviewer signature field. If the result contains only the initial entry record, the review status is not supported by the retrieved material.
- Treating a pending handover request as confirmed. The input is the draft sentence above. The comparison is between that draft and the condition in the material that an unreviewed loss report may not enter settlement. If the draft carries no review record, the request remains pending verification.
- Associating knowledge bases that use different vector models. The input is the vector model identifier of each base. The comparison is whether the two identifiers match. If they do not match, the base is excluded, and no runtime outcome of a joint association is inferred.
How to verify the setup
- Input the hypothetical material text and the draft sentence, then compare whether the retrieved content contains the review-and-confirm field. The check is satisfied when that field appears and corresponds to the same loss report. If only the initial entry record appears, record the gap.
- Input the same question through semantic retrieval and through full-text retrieval, then compare the recalled content. Observe which clauses are covered by exact term hits and which by semantic hits. No guarantee is made that the first item is a hit or that all relevant items are recalled.
- Input a paper question that clearly has no corresponding material, then compare whether the output reference array is empty. Check whether the downstream design can accept an empty result. This page does not claim that a query against an invented rule necessarily returns an empty array.