What the documents look like
Fruit-store procurement records revolve around purchase orders, acceptance sheets, pesticide-residue test reports and loss records. Wholesale market quotations and supplier quotes change daily, acceptance grades and loss entries arise per batch, and residue reports accompany the goods. For operations-permission and handover work, these materials require three separate facts to be kept apart: whether a document has been stored and trained, whether retrieval returns a match, and whether a returned match carries any authorization. The knowledge-base search node returns cited content as an array whose length can be zero, so the downstream chain still runs when nothing is found. Permission decisions, handover confirmation and business execution sit outside retrieval. Retrieval does not grant authority.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; check the vector-model identifiers and exclude those that do not match. Search mode can be semantic, full-text or hybrid. Hybrid search merges multiple result paths through RRF, and RRF is a fusion process rather than a separate switch. Minimum relevance is a filter threshold that takes effect only under semantic search or when result reranking is used; it is not a permission or entity filter, and it does not exclude unrelated businesses or unauthorized records. The citation limit is measured in tokens and caps how much content a single search may cite. Question optimization performs coreference resolution and question expansion and adds one model call. Image retrieval requires a corresponding image input and model capability, and reranking requires a rerank model enabled by configuration. Optional paths shown in the documented flow diagram do not mean every request executes all of them.
Hypothetical paper exercise: a handover request states that the fresh-produce supervisor of Store A transfers, on 10 March 2026, the review permission for the 9 March acceptance sheet (batch B-0309, quantity 40 boxes) to Buyer B, with handover status "pending receipt". The source wording available for comparison is: "Acceptance sheet B-0309 records 40 boxes arriving on 9 March; the acceptance conclusion is hold pending inspection." "Pesticide-residue report R-0309 states that the test result for this batch is not yet issued." "Loss record L-0309 records 2 boxes lost that day, cause undetermined." The draft sentence to be judged is: "This batch passed acceptance and the loss cause is confirmed." In this sample the acceptance conclusion is hold pending inspection, the test result is not yet issued, and the loss cause is undetermined, so the draft sentence does not match the materials.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the procurement and acceptance document bases, after confirming their vector-model identifiers match | A shared vector model is the precondition for association; check each identifier and exclude mismatches |
| Search mode | First test the same question with semantic search and full-text search separately | Compare how each mode matches an acceptance-sheet number and a natural-language question |
| Citation limit | Set 2000 tokens as an exercise starting point | Compare whether cited content stays complete across long and short acceptance sheets and test reports, then adjust |
| Minimum relevance | Set 0.5 as an exercise starting point | Label relevant and irrelevant samples first, then observe which cited content survives the filter |
| Question optimization | Enable in multi-turn follow-up scenarios, disable for single clear questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when a rerank model is available and observe the 0-1 score | Compare candidate ordering before and after reranking and filter low-relevance results by score |
These values are hypothetical exercise starting points, not official defaults.
Three common errors
- Treating a retrieval match as a permission confirmation. In the sample, the handover request is pending receipt; retrieval may return cited content for that request, yet the pending-receipt status does not mean permission has transferred. Check: input the handover request text, compare the retrieval output with the request's status field, and observe whether the output contains the pending-receipt wording, confirming that the retrieval result does not change the authorization state.
- Treating an unstated test conclusion as a pass. In the sample, R-0309 states the test result is not yet issued, while the draft sentence says acceptance passed. Check: input the test report text and the draft sentence, compare the test-conclusion field with the draft wording, observe whether the two agree, and record the unsupported wording.
- Associating bases that use different vector models. Check: input the list of vector-model identifiers for the bases to be associated, compare whether the identifiers agree, observe whether any mismatch exists, exclude the mismatching ones and record this, and do not run cross-model retrieval.
How to verify the setup
- For the draft sentence "This batch passed acceptance and the loss cause is confirmed" in the hypothetical sample, input the source text of acceptance sheet B-0309, test report R-0309 and loss record L-0309, compare the acceptance conclusion, test result and loss cause fields, and observe whether the draft sentence matches the source text. The expected phenomenon is that the draft sentence does not match the source text and counts as unsupported wording.
- For the handover request in the hypothetical sample, input the request text, compare the handover-status field with the pending-receipt wording, and observe whether the retrieval output contains that status. The expected phenomenon is that the output contains pending receipt, which is judged as pending verification and does not support a conclusion that permission has transferred.
- For batch B-0309 in the hypothetical sample, input the batch number and test it with semantic search and full-text search separately, compare the cited content from the two modes, and observe how a number-type query matches under full-text search. The expected phenomenon is that the two modes differ; record the actual observation without presuming that either mode must match.
For empty-result downstream handling, a clearly stated paper input of [] can be used to check the downstream design; this does not assert that a fictional rule search returns []. Missing evidence stays missing, pending receipt does not mean received, and an unissued test result does not mean passed.