What the documents look like
Convenience-store fresh food is ordered at store level against short shelf life and high frequency. Store staff work from fresh-food order forms, damage write-off forms, near-expiry handling sheets and cold-chain temperature records. These materials draw on POS sales, the head-office ordering system, store inventory and cold-chain temperature logs, and they update by day and by shift. The update rhythm matters here because retrieval returns whatever version has been ingested, so a search run during a shift can surface an earlier version of the same form.
The people who search these materials in store service and internal training are store managers, fresh-food buyers and regional supervisors. They use retrieval to check operating wording against ingested documents. The retrieval result presents the content of ingested materials only. It does not approve an order, confirm a write-off or release a temperature hold.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases on different vector models cannot be linked together. Before a search, check the knowledge base identifier and the vector model identifier, exclude those that do not match, and record what was observed. Do not run a cross-model search to compare missed recalls.
Hypothetical drill sample: a store's early-shift fresh-food order form for 10 March 2025 reads "Store number A-017, order date 10 March 2025, rice ball orders 12 cases, sandwich orders 8 cases, submission status pending regional supervisor review." The near-expiry handling sheet reads "Store number A-017, date 10 March 2025, sandwich near-expiry quantity 3 cases, handling method pending store manager confirmation." A proposed draft sentence reads "This store completed sandwich handling under the write-off process on 10 March 2025." Comparing the draft against both source texts on the same subject, same date and same field: the order form states only that submission is pending review, and the handling sheet states only that the handling method is pending store manager confirmation. The fact of completed write-off is not stated in the sample, so it is unsupported.
Search mode is compared across semantic, full-text and hybrid retrieval for the test question. No single mode is assumed to suit all materials. The citation limit is measured in tokens; test content completeness with materials of different lengths. The minimum relevance threshold takes effect only in semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered. Image retrieval requires a corresponding image input and model capability, and reranking requires a model enabled under its settings. The optional paths shown in the overall flow diagram do not mean every request executes all of them. Cited content is output as an array whose length can be 0, and the empty-result path still runs.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the same-vector-model base containing fresh-food order forms, damage write-off forms, near-expiry handling sheets and temperature records | Check the base identifier and vector model identifier first, compare against the base holding the material to be searched, exclude mismatches |
| Search mode | Run the sample question under semantic retrieval first, then run exact terms such as "A-017" and "pending review" under full-text retrieval | Compare whether returned content covers the sample source texts under each mode, record differences |
| Citation limit | Hypothetical starting point set to the token amount needed to cover both sample source texts | Test content completeness with materials of different lengths, compare against sample source text length before adjusting |
| Minimum relevance | Label relevant and irrelevant passages in the sample first, then set a threshold | Takes effect only in semantic retrieval or result reranking; compare filtering outcome against labelled results |
| Question optimization | Enable for short follow-ups such as "what about the sandwiches", disable for complete questions | Compare original and optimized questions, check whether subject and object were altered |
| Result reranking | Enable when candidates are numerous and a rerank model is configured | Compare ordering before and after reranking, check scores against the sample question |
Three common errors
- Reading "submission status pending regional supervisor review" as review passed. Input is the sample order form text; comparison is between the values "pending review" and "reviewed"; the observable result is that the source text shows only pending review, and review passed is not stated. Record it as pending verification rather than writing it as approved.
- Reading "handling method pending store manager confirmation" as write-off completed. Input is the sample near-expiry handling sheet text; comparison is between the values "pending confirmation" and "completed"; the observable result is that the source text shows only pending confirmation, and write-off completion is not stated. Keep the evidence gap open.
- Treating a similarity score as a permission or entity filter. Input is the returned score and the knowledge base identifier; comparison is between the score level and the authorization scope of store number A-017; the observable result is that the score carries no permission information and retrieval does not constitute authorization. Connect a separate business authorization step.
How to verify the setup
- Input the sample order form text and the draft sentence "This store completed sandwich handling under the write-off process on 10 March 2025"; compare whether the draft matches the source text on the same subject, same date and same field; the observable result is that the completed write-off stated in the draft is not present in the source text, so it is judged unsupported.
- Input the sample near-expiry handling sheet text and the handover request "please continue downstream operations on the basis of completed write-off"; compare the status field the request relies on; the observable result is that the source text states only pending store manager confirmation, the completion status the request relies on is absent, so it is judged pending verification.
- Input an explicit paper empty-result array []; check whether the downstream design still runs the path when the citation length is 0; the observable result is that the path continues, without claiming that a fabricated rule search necessarily returns [].
Retrieval checks record observations and failure branches only. They do not guarantee a first-hit result, full recall or that an unrelated query returns nothing. For cross-model bases, check identifiers first and exclude mismatches. Pending review does not demonstrate absence of operating authority; signing, reaching cooperation and contract effectiveness are three distinct facts. Missing evidence stays missing.