What the documents look like
Supermarket supplier accounts settle through purchase-and-sale or consignment payables. The supplier does not necessarily operate a staffed counter on site, which separates this category from concession counters in large stores. The documents handled day to day are procurement contracts, delivery acceptance notes, rebate agreements and reconciliation statements. They arrive from ERP purchasing, supplier portals, POS sales and finance systems, and they refresh mainly by week and by month.
For product and marketing copy drafting, these documents supply the checkable facts. A draft sentence may state a settlement method, a basis for accepted quantities, or a rebate percentage. The retrieval scope follows from that: the knowledge base should cover these document types, and retrieval results are used to support specific statements in the copy. The task on this page is limited to retrieval and recall for copy drafting. Whether a particular promotional sentence is supported by the material is a separate question from enterprise identity verification.
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 with different vector models cannot be associated together. Check the vector model identifier, exclude those that do not match, and record what is actually observed. Do not run a cross-model retrieval to compare missed recalls.
Retrieval searches the whole knowledge base. A collection is only for classifying data and does not affect search results. One retrieval combines question optimization, semantic search, full-text search, hybrid search, image search, RRF fusion and reranking according to the input and search parameters. The optional paths shown in the overall flow diagram do not mean every request executes all of them. Image search requires image input and model capability. Reranking requires a model and must be enabled by configuration.
Reference content is output as an array, and its length can be zero. The output link still runs even when nothing is found. Minimum relevance is a filter threshold, effective only in semantic search or when reranking is used. It is a filter condition and does not perform permission or entity isolation. Retrieval itself does not constitute authorization.
Hypothetical paper exercise: the source text for checking reads, "Party A is XX Supermarket, Party B is YY Food Supplier. The two parties signed the 2025 annual purchase and sale contract. Settlement is monthly with 30 days, the rebate is calculated at 2% of purchase amount, and reconciliation is due before the 5th of each month." The draft sentence to be judged reads, "After entering the store, this supplier settles weekly, and the rebate ratio is 5%." The check compares subject, period and field. The source states monthly settlement with 30 days and 2%. The draft sentence states weekly settlement and 5%. The same field carries mutually exclusive values, so the sentence conflicts with the material and is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the supplier settlement base that uses the same vector model | Compare vector model identifiers across bases, associate matching ones, exclude non-matching ones |
| Search mode | Test semantic search first, then full-text search on the same question | Compare the returned content of the two modes using one test question |
| Reference limit | Set to 800 tokens for this exercise | This is a starting point for this exercise, not an official default. Compare with documents of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Set to 0.4 for this exercise | This is a starting point for this exercise, not an official default. Label relevant and irrelevant samples first, then compare the filtering effect |
| Question optimization | Enable in multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject was altered |
| Reranking | Enable by configuration when a rerank model is available | Compare ordering before and after enabling, and confirm how question completeness affects it |
Three common errors
- Treating a retrieval hit as an authorization basis. Input is a retrieved rebate agreement fragment. The comparison object is whether that fragment states the authorization for the current subject and period. If the fragment only describes evidence at a related stage, the judgment is that support is absent, and settlement or final copy approval is not performed on that basis.
- Treating the similarity threshold as an entity or permission filter. Input is a low-relevance query and an unrelated business record. The comparison object is the threshold setting and the record's ownership. If the record is still recalled, the threshold does not perform isolation, and a separate permission implementation is required.
- Associating knowledge bases across vector models. Input is two bases with different vector model identifiers. The comparison object is the two model identifiers. If they do not match, exclude them, record the observation only, and do not assert any runtime result of associating them.
How to verify the setup
- Input the source text and the draft sentence from the hypothetical exercise. Compare subject, period and field values. If the same field carries mutually exclusive values, the judgment is a conflict, and the draft sentence is not supported.
- Input an explicit paper-based empty result array []. Compare how the downstream node is designed to handle that input, and check whether the empty-result link runs as designed. Do not claim that a fictional rule retrieval necessarily returns an empty array.
- Input the same test question through semantic search and full-text search. Compare the differences in returned content. If an identifier-type query is matched more precisely under full-text search, record that observation. Do not guarantee a first hit or complete recall.