What the documents look like
A hypermarket fresh-food concession counter is typically operated on site by a supplier, with revenue shared between the supplier and the store. Order fulfilment and after-sales answers therefore involve three parties at once: the store's fresh-food department, the concession supplier, and the store manager. The documents that matter most are the concession contract, the daily sales report, the shrinkage record, and the food inbound inspection record. Data sources include the POS system, supplier delivery notes, the food-safety ledger, and store stock counts. These materials update daily or per batch, so reconciliation work compares sales flows, delivery batches, and shrinkage records inside one shared date range.
For retrieval, the input is usually fragments of these documents, and the output is an array of cited content whose length can be zero. That output does not carry any authorization judgement. A concession contract fragment may state a liability holder, a sales-share rate, a counter number, a product category, and an effective date range. A daily sales report may state a sales amount and a shrinkage quantity for the same counter and the same date. A shrinkage record may carry a status such as pending review. A food inbound inspection record may carry a batch identifier and a delivery date. The retrieval step returns candidate text; it does not decide who owes what.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Before association, the vector model identifier of each candidate library is checked, non-matching libraries are excluded, and the remaining association is recorded. Search mode can be configured as semantic search, full-text search, or hybrid search. Hybrid search merges multiple recall paths through RRF; RRF is a fusion process inside hybrid search, not a separate switch. The minimum relevance value takes effect only under semantic search or when result re-ranking is used. That threshold filters low-relevance results and does not perform permission, entity, or business-scope isolation. The citation limit is measured in tokens and controls how much content a single search may cite. Question optimization performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires image input and matching model capability. Re-ranking requires an available model and must be enabled in settings. The optional paths listed in the overall retrieval flow do not all execute on every request.
Hypothetical paper test case: a concession contract clause reads, "Party B operates leafy greens at counter 3; shrinkage is borne by Party B; Party A collects a share of eight percent of sales," with an effective range of 1 January 2025 to 31 December 2025. A daily sales report shows a leafy-green sales amount and shrinkage quantity for the same counter and the same date. A draft sentence under review reads, "Shrinkage at this counter is borne by Party A." After retrieving the contract fragment, the liability-holder field of the draft is compared with the contract text: the contract names Party B, the draft names Party A, and the same field holds mutually exclusive values, which is a contradiction. If retrieval does not return that contract fragment, the result is recorded as unsupported, not as a denial.
Configuration and implementation checks
The table covers only settings or implementation checks that the official documents describe and that relate to this feature and this task. The values are starting points for this hypothetical trial and are not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the library containing this concession contract | Check that the vector model identifier matches before associating; exclude libraries with a different model |
| Search mode | Hybrid search | The contract carries exact fields such as party names and dates; hybrid search uses semantic and full-text recall together, then fuses through RRF |
| Citation limit | Set a token amount that covers the liability clause and the date range | Test with materials of different lengths to see whether content stays complete, then adjust the limit |
| Minimum relevance | Label relevant and irrelevant samples first, then set the threshold | This value takes effect only under semantic search or result re-ranking and does not act as a permission filter |
| Question optimization | Enable for follow-up questions that contain references | Compare the original question with the optimized question and check whether the subject was altered |
| Result re-ranking | Enable in settings and confirm model availability | The re-rank score is used for filtering and does not replace manual comparison against the original text |
Three common errors
- Treating the minimum relevance value as a permission or entity filter. Input a query unrelated to the concession contract and observe whether returned results can still include content from other counters. A relevance score does not enforce entity isolation; access scope is checked separately.
- Forcing association of libraries built on different vector models. Input two libraries with different vector model identifiers, compare the identifiers, and when they differ, exclude that library and record the exclusion. Do not run cross-model retrieval to compare missed recalls.
- Treating a handover request marked pending review as already approved and executed. Input a shrinkage-allocation request labelled "pending review" and compare its status field. Pending review does not show that operation permission is absent, and it is not an approval. The evidence gap is preserved.
How to verify the setup
- Input the contract text and the draft sentence from the hypothetical sample above, and compare the liability-holder field. If retrieval returns the contract fragment and the draft subject is mutually exclusive with the contract text, the contradiction condition is met. If the fragment is not returned, record the result as unsupported.
- Input a follow-up question that contains a reference, and compare the retrieval terms before and after question optimization. If the subject after optimization matches the original text, the coreference-resolution condition is met. If the subject was altered, record that optimization result for review.
- Input an explicit paper empty result
[]and check whether the downstream design accepts cited content of length zero. The empty-result handoff can be checked with that input. Do not claim that a fabricated rule search necessarily returns[].