What the documents look like
Fresh-food supplier claims pursue responsibility for the quality of goods on arrival. They do not address natural in-store shrinkage, so the retrieval material is built around per-batch quality inspection and receiving records, with monthly reconciliation providing the summary layer. The documents handled day to day include quality exception forms, photo evidence, deduction notices and claim reconciliation statements. These four material types do not share one field structure: a quality exception form records batch and determination, photo evidence records the condition of goods, a deduction notice records amount and basis, and a claim reconciliation statement records the conclusion of the exchange. Product and selection questions must be able to answer which material corresponds to a given batch and which contractual clause supports a given deduction, so the retrieval step has to bring these materials back accurately.
The category also differs from near-expiry discount handling, which works on a different responsibility basis and a different document set. For this page, the relevant consequence of a retrieval error is a supplier dispute, an absorbed cost of bad goods, or an interruption in supply.
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. Recall paths include semantic search, full-text search and hybrid search; image description retrieval or image vector retrieval requires the corresponding image input and model capability, and reranking requires a rerank model enabled in the settings. These paths do not all execute on every request. Hybrid search merges multiple result paths through RRF, and RRF is a fusion process. The minimum relevance value takes effect only in semantic search or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The reference limit is measured in tokens. Question optimization performs coreference resolution and question expansion.
A search hit is not an authorization. Retrieval does not decide whether a deduction is effective, whether a supplier has agreed to pay, or whether a claim has been approved.
Hypothetical paper test sample. The source text to be checked: "Quality exception form QA-0417 states: batch A-0417, arrival date 17 April, leafy vegetables, receiving determination failed, responsibility attributed to the supplier." "Deduction notice K-0417 states: under contract appendix three, clause 2, a deduction applies to batch A-0417, amount pending reconciliation confirmation." The draft sentence to be judged: "The claim for batch A-0417 is already effective and the supplier has agreed to pay." The subject, object, status and validity scope needed for the judgment are present in the source text: the subject is the supplier, the object is batch A-0417, and the status is that the deduction amount is pending reconciliation confirmation. The draft sentence turns a pending confirmation into an effective claim and an agreement to pay, which the source text does not support.
Configuration and implementation checks
| Configuration item | Setting on this page | Basis and checking method |
|---|---|---|
| Associated knowledge bases | One knowledge base each for quality exception forms and deduction notices; first check whether the vector model identifiers match | Only the same vector model may be associated together; exclude and record any base whose identifier does not match |
| Search mode | Hybrid search as the hypothetical starting point for this trial | Batch numbers and contract clauses are exact queries, so results must be compared with pure semantic search |
| Reference limit | Set to 2000 tokens as the hypothetical starting point for this trial | Test with quality exception forms of differing length to see whether content is complete, then compare with an adjacent value |
| Minimum relevance | Label relevant and irrelevant samples first, then set the threshold | Takes effect only in semantic search or result reranking |
| Question optimization | Enabled for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether the subject has been altered incorrectly |
| Empty-result handling | Check the downstream design against a paper input of [] | Reference content may have length 0, and the output path still runs through |
Three common errors
- Treating a retrieval hit as authorization. Input the query "batch A-0417 deduction notice," compare the retrieval result with the authorization status, and note that treating recalled content as an effective claim is an error. Retrieval does not perform authorization judgment.
- Writing a pending status as effective. Input the draft sentence and the source text, compare the status field, and note that the source states the amount is pending reconciliation confirmation while the draft says the supplier has agreed to pay. The draft is unsupported.
- Forcing an association across vector models. Input the vector model identifiers of the two knowledge bases, compare whether they match, and note that associating them when the identifiers differ is an error. Check only the identifiers and exclude the non-matching base.
How to verify the setup
- Input the source text of quality exception form QA-0417 together with the draft sentence, compare subject, object and status, and require that the draft sentence be judged unsupported with the evidence gap preserved.
- Input batch number A-0417 as a query, compare the recalled content from hybrid search with that from pure semantic search, and observe whether the exact number is consistently hit. When it is not hit, record the failure branch; a first-position hit is not guaranteed.
- Input a paper empty result of [], compare whether the downstream node still runs through, and require that the output path runs through. Do not claim that a fictional rule makes retrieval always return [].