What the documents look like
Fresh-food concession counters in hypermarkets are frequently staffed on site by suppliers who operate the counter and share sales revenue with the store. The documents handled day to day are the concession contract, daily sales reports, loss sheets and food purchase inspection records. These materials originate from the POS system, supplier delivery notes, food safety ledgers and store stock counts, and they are updated daily or per batch. Product information and selection questions draw on these documents to retrieve specifications, settlement terms, loss ownership and inspection status. Retrieval returns candidate passages only. It does not confirm authorization, and the similarity score does not separate business entities or filter records by permission.
A retrieval step that misreads the loss owner or the inspection state feeds directly into food safety penalties, gross-margin disputes and customer claims. The subject, object, status and validity range recorded in each document are the basis for any later judgment. The search node can be linked to one or more knowledge bases that use the same vector model; bases built on different vector models cannot be linked together.
Constraints at the retrieval step
The search mode can be semantic retrieval, full-text retrieval or hybrid retrieval. Hybrid retrieval merges multiple result channels through RRF, and ReRank performs a second ordering of candidate text. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. The citation limit is measured in tokens. Quoted content is output as an array whose length can be zero, so the output path still runs when nothing is found. Question optimization performs coreference removal and question expansion and adds one model call before the actual search.
Hypothetical paper exercise sample. A concession contract text states: "Party A is the fresh-food department of XX hypermarket; Party B is the on-site supplier; counter sales are shared daily; loss is borne by Party B; the contract is valid from 1 March 2025 to 28 February 2026." A food purchase inspection record states: "10 April 2025, batch A-0410, inspection passed." A draft sentence to be assessed reads: "The supplier's April 2025 loss is borne by Party A." The check compares the loss-ownership field in the draft sentence with the same field in the contract text. The contract states that loss is borne by Party B, so the two values in the same field are mutually exclusive and the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Explanation |
|---|---|---|
| Linked knowledge bases | Select knowledge bases that use the same vector model | Bases on different vector models cannot be linked; verify identifiers first and exclude those that do not match |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | No single mode is assumed to suit all document types |
| Citation limit | Starting value for this hypothetical trial set at 2000 tokens | Chosen because concession contract passages are long; compare against 1000 tokens for content completeness |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Takes effect only under semantic retrieval or result re-ranking |
| Question optimization | Compare the original question with the optimized question | Check whether the subject has been altered; first-hit ranking is not guaranteed |
| Result re-ranking | Enabled when a model is available and configured | An optional path; it does not mean every request executes it |
Three common errors
- Treating a retrieval score as a permission filter. Input is an ordinary product question; compare the retrieved content with the range the current user may access; observe whether records outside that range appear. The relevance score does not perform entity isolation. The same-vector-model requirement for linked bases does not extend to external systems such as business registration sites.
- Treating a pending-review state as approved execution. Input is a loss-allocation sheet under review; compare its status field with the value "approved"; observe whether it is treated as a completed allocation. Pending review does not show that an operating permission is missing, and it is not approval.
- Treating a quotation validity period as a supply term. Input is a supplier quotation; compare the quotation validity field with the contract supply-term field; observe whether the two are mixed. These are different fields, and a quotation validity period does not indicate a supply term.
How to verify the setup
- Input the contract text from the hypothetical sample and the draft sentence "The supplier's April 2025 loss is borne by Party A." Compare the loss-ownership field. If the contract states that loss is borne by Party B, the draft sentence is unsupported and the evidence gap must be recorded.
- Input a loss sheet under review. Compare its status field with the value "approved." If the status is pending review, a completed allocation cannot be determined, and the conclusion remains pending verification.
- Input an empty quoted-content array []. Compare whether the downstream design accepts an output of length zero. If the downstream does not define handling for an empty result, record that gap. Do not claim that retrieval certainly returns an empty array.