What the documents look like
Convenience-store fresh food is ordered by the store on a short shelf-life, high-frequency basis, and it is not handled as a joint-operation revenue share. The people involved are store managers, fresh-food buyers and regional supervisors. The working materials are fresh-food order forms, damage and loss reports, near-expiry handling sheets and temperature records. These materials update daily or by shift, and the consequences they relate to are sales of expired goods, lost sales from stockouts and food-safety complaints.
Operating reports and exception reviews need the order, loss, near-expiry and temperature materials for one day or one shift to sit inside the same retrieval scope for cross-checking. The retrieval step itself returns quoted content only. It does not carry authorization, approval or database writes.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be linked together; the identifiers must be checked first and non-matching ones excluded.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. Result re-ranking is performed by a ReRank model as a second ordering pass. The citation limit is counted in tokens. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval. Quoted content is output as an array, and its length can be zero, so the downstream path still runs when nothing is found.
Image retrieval requires a corresponding image input and model capability. Re-ranking requires the model to be available and enabled according to configuration. These optional paths do not mean every request executes all of them.
Hypothetical paper exercise sample: the source material reads "The 10 March 2025 fresh-food order form for Store A (number XZ-0310) states: sandwiches ordered 40, received 40. The same store's loss report for the same day (number BS-0310) states: sandwiches lost 6, reason near expiry." The draft sentence to be judged is "On 10 March 2025, Store A ordered 40 sandwiches and all 40 were sold." That draft sentence lacks a sales field, and placed alongside the recorded loss of 6, the existing material cannot support the claim that all were sold.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the same-vector-model knowledge base holding fresh-food order forms, loss reports, near-expiry handling sheets and temperature records | Check whether the vector model identifiers of each knowledge base match, and exclude those that do not |
| Search mode | Run the sample question once with semantic retrieval and once with full-text retrieval | Compare whether the quoted content from each mode covers the order form and the loss report text |
| Citation limit | Hypothetical starting point of 800 tokens | Because the sample material contains several short forms, compare whether quoted content is complete at different limits |
| Minimum relevance | Hypothetical starting point of 0.3 | First mark relevant and irrelevant passages in the sample by hand, then compare recall differences at different thresholds |
| Question optimization | The sample question is complete on its own, so leave it off first | Compare whether the optimized question changes the sample subject when on versus off |
| Result re-ranking | Leave it off when candidates are few | Compare ordering before and after enabling, and confirm the ReRank model is available and enabled according to configuration |
Three common errors
- Treating retrieval recall as authorized execution. After the sample question is entered, the system returns only an array of quoted content. That does not mean the order form was approved, the loss report was confirmed or the temperature record was synchronized. During checking, compare the output type against business status fields and observe whether any authorization-type conclusion appears.
- Treating minimum relevance as an entity or permission filter. That value is a 0-1 filtering threshold and takes effect only under semantic retrieval or when result re-ranking is used. During checking, compare recalled content at different thresholds and observe whether relevant material inside the same-vector-model knowledge base is wrongly excluded.
- Forcing knowledge bases with different vector models to be linked together. The documentation states that only knowledge bases with the same vector model can be linked. During checking, inspect each knowledge base identifier first, exclude those that do not match, and record the actual observation without inferring what would happen after linking.
How to verify the setup
- Enter the sample question and the source material text above, and compare whether the quoted content contains both order form number XZ-0310 and loss report number BS-0310. If only one of them is recalled, record the gap and check the linked knowledge base scope.
- Enter the draft sentence to be judged, "On 10 March 2025, Store A ordered 40 sandwiches and all 40 were sold," and compare whether a sales field exists in the recalled content. If only order and loss fields are present, the sentence is unsupported and the evidence gap remains.
- Enter a paper question that clearly has no corresponding material, and check whether the quoted content is an array of length zero while the downstream path still runs. This observation shows only that empty-result handling works; it does not mean that a query against a fictional rule necessarily returns an empty array.