What the documents look like
A fresh-food concession counter in a hypermarket is commonly operated on site by a supplier, with revenue shared between the store and that supplier according to sales. The material that retrieval has to work with therefore comes from several directions at once: the hypermarket fresh-food department, the concession supplier, and the store manager. The recurring document set is the concession contract, the daily sales report, the shrinkage record, and the food purchase inspection record. Source systems named in the brief are the POS system, supplier delivery notes, the food safety ledger, and store stocktaking. Update frequency is daily and per batch.
For cross-border and multilingual work, the same batch of material may appear in more than one language. A single retrieval request can therefore need to reach contract clauses, sales figures, and inspection evidence that were written in different languages and stored as separate documents. The brief does not state which languages, which translation workflow, or which jurisdictions are involved, so those remain open.
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 built on different vector models cannot be linked together. In practice this means the vector model identifier is checked first, and any knowledge base that does not match is excluded before any search is run. What is recorded is the identifier comparison and the exclusion, not a claim about what a cross-model search would have returned.
Semantic retrieval computes similarity from vector distance and supports cross-language understanding, so a question written in one language can match a knowledge point stored in another. Full-text retrieval depends on keyword hits and suits identifiers, model numbers, proper nouns, and error codes. Hybrid retrieval uses both and merges the two result sets through RRF. The minimum relevance value is a filter threshold that applies only in semantic retrieval or when result reranking is used. It is not a permission filter and does not isolate entities or businesses. Retrieval itself is not authorization.
The search node outputs cited content as an array, and the length of that array can be zero. A search that finds nothing still completes the output path.
Hypothetical paper exercise. The material to be checked is a concession contract whose text reads: "Party A is the XX hypermarket fresh-food department; Party B is supplier YY; the counter is located in fresh-food zone A3; the contract is valid from 1 January 2025 to 31 December 2025; the sales share is settled monthly." A draft sentence submitted for review reads: "This contract expired and terminated on 31 December 2025." The source text states only the end date of the validity period. It does not state whether termination has been completed, whether a renewal was signed, or whether any termination confirmation step occurred. The draft sentence is therefore unsupported and cannot be judged a completed termination.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the concession contract knowledge base that uses the same vector model | Compare vector model identifiers; link only when they match, exclude those that do not |
| Search mode | Start with hybrid retrieval | Contract clauses contain both natural language and identifiers; compare semantic, full-text, and hybrid recall on the same question |
| Citation limit | Hypothetical starting value of 2000 tokens | Test with contracts and shrinkage records of different lengths; compare content completeness against a shorter limit |
| Minimum relevance | Hypothetical starting value of 0.3 | Label relevant and irrelevant samples first, then compare filtering behaviour at different thresholds |
| Question optimization | Enabled for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled when the candidate set is large | Compare ordering before and after reranking, and confirm the rerank score can be used for filtering |
Three common errors
- Treating a retrieval result as an authorization record. A search that returns a revenue-share clause from a concession contract does not mean the supplier has been approved to act on that clause. Authorization status has to be checked separately against signing and effectiveness records.
- Linking knowledge bases across different vector models. When two knowledge base identifiers are submitted for a single search, comparing the vector model identifiers shows the mismatch. The mismatched knowledge base is excluded, and the observation is recorded. No cross-model search is executed to compare missed recall.
- Treating a pending review state as an approval. A shrinkage record submitted with a pending-review status is compared on its status field. Pending review does not demonstrate that an operator lacks permission, and it is not an approval.
How to verify the setup
- Input the hypothetical contract text above together with the draft sentence. Compare whether the draft sentence is supported by the source. The source states only the validity end date and does not state a completed termination. The draft sentence must be judged unsupported and must not be judged a completed termination.
- Input two knowledge base identifiers that use the same vector model. Compare whether the identifiers match. Linking is permitted only when they match. When they do not match, exclude the non-matching knowledge base and record the observation. Do not run a cross-model search to compare missed recall.
- Input an empty-citation scenario. Compare whether the output path completes. Cited content is output as an array and its length can be zero. A paper input of an empty array can be used to check the downstream design for an empty result. Do not claim that a fictional rule search necessarily returns an empty array.
The optional paths shown in the overall retrieval flow are not all executed on every request. Image retrieval requires an image input and matching model capability. Reranking requires a rerank model and must be enabled by configuration. Knowledge bases built on different vector models cannot be linked together; the check is the identifier comparison, the exclusion of non-matching bases, and a record of what was observed.