What the documents look like
Fresh-food supplier claims pursue responsibility for quality at arrival, which is a different business scope from in-store natural shrinkage handling. The main data sources are the quality inspection system, WMS acceptance, supplier contracts and financial reconciliation. Day-to-day materials include quality exception forms, photo evidence, deduction notices and claim reconciliation statements. The operating review and exception review stage retrieves batch numbers, responsibility attribution, deduction amounts and negotiation status from these materials so that procurement, quality inspection and supplier business staff can check them.
Materials arrive on two rhythms: per batch and monthly. Each batch produces quality exception forms and photo evidence; each month produces deduction notices and claim reconciliation statements. Retrieval and recall operate on materials already ingested and do not change their business status.
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; the vector model identifier must be checked first and non-matching bases excluded. Cross-model retrieval is not run to compare missed recall.
Hypothetical drill sample: a fresh-food arrival acceptance record states "12 March 2025, batch A, acceptance found the wilting ratio of leafy vegetables exceeded the contract terms, quality inspection has issued a quality exception form, photo evidence number P-0312, awaiting supplier confirmation." A separate deduction notice draft states "proposed deduction for batch A, amount to be determined after quality inspection review." A draft sentence to be judged reads "The batch A claim is established and the deduction has been executed."
This sample is used to check whether retrieval results and the draft sentence agree. The retrieval input is "batch A quality exception form deduction status"; the comparison is whether the recalled content contains the acceptance record text above, and the observation is whether the quoted content shows "awaiting supplier confirmation" and "amount to be determined after quality inspection review." The draft sentence writes a pending-confirmation state as established and a pending-review amount as executed, which does not match the status fields in the source text.
Search modes are compared across semantic, full-text and hybrid retrieval for the test question; no single mode is assumed to suit all materials. The citation limit uses the token unit; test content completeness with materials of different lengths before adjusting the limit. Minimum relevance takes effect only in semantic retrieval or when result reranking is used; label relevant and irrelevant samples first, then adjust the threshold. Question optimization performs coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered.
Quoted content is output as an array and its length can be 0. For the empty-result path, a clearly paper-based input of [] can be used to check downstream design; no claim is made that a fictional-rule search must return []. Retrieval does not handle permission or entity isolation; semantic recall and relevance do not perform authorization decisions.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the base holding quality exception form and photo evidence indexes | Check whether the vector model identifiers of each base match; exclude non-matching ones |
| Search mode | Hybrid retrieval | Use full-text retrieval for batch numbers and semantic retrieval for responsibility descriptions; compare the two recall paths |
| Citation limit | Set to a length covering one quality exception form plus photo description | Test with materials of different lengths whether quoted content is truncated |
| Minimum relevance | Value chosen after labelling relevant and irrelevant samples | Compare whether irrelevant fragments are filtered at different values; record the observation |
| Question optimization | Enabled for multi-turn follow-ups containing references | Compare the subject of the original and optimized questions for consistency |
| Result reranking | Enabled when a rerank model is available and configured | Compare candidate ordering before and after reranking; record the actual observation |
The values above are hypothetical trial starting points for this page, not official defaults. Reranking requires a model and must be enabled by configuration; image retrieval requires corresponding image input and model capability. The optional paths shown in the overall document flow diagram do not mean every request executes all of them.
Three common errors
- Treating a quality exception form awaiting supplier confirmation as an established claim. Input is the acceptance record text above; compare the "awaiting supplier confirmation" field; observe whether the draft sentence rewrites that state as established. Pending review does not demonstrate a lack of operating permission; signing, cooperation reached and contract effectiveness are three distinct facts.
- Treating a deduction notice draft with a pending amount as an executed deduction. Input is the deduction notice draft text; compare the "amount to be determined after quality inspection review" field; observe whether the draft sentence writes the proposed deduction as executed. A quotation validity period does not indicate a supply period; a field not stated, or supported only by a related-stage piece of evidence, is unsupported or pending verification.
- Associating knowledge bases with different vector models and then inferring retrieval behaviour. Input is the vector model identifier of each base; compare whether the identifiers are the same; observe whether non-matching ones were excluded. Do not invent run results such as missed recall or redundant returns after joint association; only check identifiers, exclude non-matching ones and record the actual observation.
How to verify the setup
- Input the acceptance record text above and the retrieval term "batch A quality exception form deduction status"; compare whether the recalled content contains the source text; observe whether the quoted content shows "awaiting supplier confirmation" and "amount to be determined after quality inspection review." The expectation is that recalled content matches the status fields in the source text; first-hit ranking or full recall is not guaranteed.
- Input the draft sentence "The batch A claim is established and the deduction has been executed"; compare the draft sentence with the status fields in the source text; observe whether the two mismatches of pending confirmation and pending review are identified. The expectation is that the draft sentence is judged unsupported; no claim is made that this was actually tested.
- Input the vector model identifiers of the knowledge bases; compare whether the identifiers are identical; observe whether the joint-association condition is met. The expectation is that non-matching identifiers are excluded and the actual observation is recorded; a missing model identifier is recorded as unknown, and retrieval results are not used to infer authorization or effectiveness status.