What the documents look like
Convenience-store fresh food is ordered by the store at high frequency against short shelf life. Ordering, waste reporting, near-expiry handling and temperature records are produced from POS sales, the headquarters ordering system, store inventory and cold-chain temperature logs. These materials update by day or by shift. Typical roles are the store manager, the fresh-food buyer and the regional supervisor. Errors show up as expired sales, stock-out losses and food-safety complaints.
For this page the retrieval step only recalls quoted content from material already stored in a knowledge base. Authorization status and handover responsibility are judged by a person reading the original material. A fresh-food order form, a waste report, a near-expiry handling sheet and a temperature record each carry different subjects and different fields. A retrieval result does not change the approval status of any document.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. The knowledge base is the smallest search unit; a collection is used only for classification and does not affect search results. Semantic search, full-text search and hybrid search are compared against the test question; no single mode is assumed to suit all fresh-food material. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Question optimization performs coreference resolution and question expansion. Hybrid search merges multiple result paths through RRF; RRF is a fusion process, not a switch. Image retrieval requires image input and the corresponding model capability. Reranking requires a model and is enabled according to settings. Retrieval and similarity scores do not constitute authorization and do not act as an entity or permission filter.
Hypothetical paper exercise sample. Source text one: "Store 07 submitted fresh-food order form XD-0310-07 on 10 March 2025, day shift, item rice ball, quantity 40 units, status pending review." Source text two: "The temperature record WD-0310-07 for the same store and date notes the refrigerated cabinet at 6 degrees Celsius, recorded by the shift employee." Draft sentence to be judged: "The order form has been approved and can be used to schedule next-day production." The subject, object, status and effective scope must be checked. The source states only pending review and does not state an approval field or approval time, so the approval conclusion in the draft is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Link only fresh-food material bases confirmed to use the same vector model | Compare against the identifier of a base using another vector model; exclude mismatched identifiers and do not link across models |
| Search mode | Try full-text search for order-form fields containing numbers; try semantic search for natural-language questions | Compare hit content against the hybrid search result for the same test question |
| Citation limit | Hypothetical starting point 1500 tokens | Compare against actual material of different lengths to see whether content is complete, then adjust from observation |
| Minimum relevance | Hypothetical starting point 0.4 | Label relevant and irrelevant samples first, then adjust after comparing against the labelled result |
| Question optimization | Enable for multi-turn follow-ups and short questions; disable for independent, clear questions | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enable when candidates are numerous and the text question is clear | Compare against the candidate order without reranking and observe changes in the relevance score |
Three common errors
- Treating a retrieval hit as authorization. Input: pending-review order form XD-0310-07 and the draft sentence. Compare the status field in the source with the conclusion in the draft. The observed phenomenon is that the source states only pending review while the draft says approved; that conclusion is unsupported.
- Treating the similarity threshold as a permission or entity filter. Input: the same question and the minimum relevance setting. Compare the recalled content against the access permission list. The observed phenomenon is that the threshold affects relevance filtering only and does not exclude records without permission.
- Writing RRF as a switch, or forcing links between bases with different vector models. Input: the vector model identifiers of two bases. Compare whether the identifiers match. The observed phenomenon is that mismatched identifiers must be excluded and recorded; their missed-recall results are not compared.
How to verify the setup
- For the hypothetical sample, input source text one, source text two and the draft sentence. Compare whether "pending review" and "approved" share the same subject, the same period and the same field. The observed phenomenon is that the source states no approval field, so the judgment is unsupported and the evidence gap is retained.
- For empty-result handling, input the paper citation content []. Compare whether the downstream design proceeds with a zero-length citation. The observed phenomenon is that the output path still runs; no assertion is made that a fictional rule must return [].
- For cross-model linking, input the vector model identifiers of two knowledge bases. Compare whether the identifiers match. The observed phenomenon is that mismatched identifiers are excluded and the actual observation is recorded; no cross-model retrieval is performed to compare missed recall.