What the documents look like
Fresh-food supplier claims pursue quality liability for goods as received. The work does not cover ordinary in-store shrinkage, which belongs to a different handling track. The materials that matter here are quality exception forms, photo evidence, deduction notices and claim reconciliation statements. These are produced from quality inspection records, WMS acceptance results, supplier contracts and financial reconciliation.
The document set has two retrieval shapes. Batch numbers, acceptance document numbers and deduction document numbers are exact identifiers, and full-text retrieval suits them. Quality exception narratives and compensation negotiation notes are natural-language content, and semantic retrieval suits them. The same claim file therefore contains both kinds of query target, and a single search mode does not fit every question in it.
Materials arrive from quality inspection, procurement and supplier-facing staff. Retrieval returns cited content only. A returned citation does not constitute authorization, approval or reconciliation confirmation, and the retrieval step does not decide whether a deduction may be executed.
Constraints at the retrieval step
The dataset search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the base, not the collection; collections classify data and do not affect search results. Before retrieval, confirm that the vector model identifiers of the candidate bases match. Bases with different identifiers cannot be associated together. At this step the work is to check identifiers and exclude non-matching bases; it does not run cross-model retrieval to compare recall.
Search mode is compared per test question across semantic, full-text and hybrid retrieval. No single mode is assumed to suit all claim materials. The citation cap is expressed in tokens, and it is not a top-k item count. The minimum relevance value applies only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval. Reranking requires an available model and must be enabled in settings. Image retrieval requires image input and the corresponding model capability.
An empty citation output has length 0, and the downstream chain still runs. The overall retrieval flow combines several optional paths, and not every path runs on every request.
Hypothetical paper exercise sample. A supplier-facing staff member submits a handover request with three source sentences. Quality exception form QZ-2026-0417 states: goods arrived on 17 April 2026, batch PC-0417, acceptance conclusion partly nonconforming. Deduction notice KK-2026-0418 states: a deduction is proposed for that batch, status pending supplier confirmation. Claim reconciliation statement DZ-2026-04 states: the batch deduction amount is included in the April reconciliation, confirmation status not stated. The draft sentence to be judged is: "The batch deduction has been confirmed by the supplier and can be executed." The handover request is: "Please retrieve the three materials above and answer whether the deduction is confirmed."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate the quality, contract and reconciliation bases | Check first that vector model identifiers match; exclude non-matching bases and do not run cross-model retrieval |
| Search mode | Trial run hybrid retrieval first | Compare cited content for the same question against semantic retrieval and full-text retrieval separately |
| Citation cap | Set a starting token amount sufficient to cover the three source sentences | This value is a trial starting point only; compare completeness across short and long materials before adjusting |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | The value applies only under semantic retrieval or reranking; compare cited items before and after filtering |
| Question optimization | Enable for the question "is the deduction confirmed" | Compare the original question with the optimized question and check whether the subject and batch were altered |
| Result reranking | Enable per settings and confirm the model is available | Compare rerank scores and vector scores for the same candidate to see ordering differences |
Three common errors
- Treating a pending status as confirmed. In the sample, the deduction notice status is pending supplier confirmation and the reconciliation confirmation status is not stated, while the draft sentence says the supplier has confirmed. Input is the three source sentences; the comparison is between the values "pending confirmation" and "confirmed"; the observable result is that the draft sentence does not match the source values.
- Treating retrieval output as execution authority. A returned citation shows that material was recalled; it does not show that a deduction was approved. Input is the handover request and the retrieved citations; the comparison is between "material was retrieved" and "execution authority exists"; the observable result is that the request contains no authorization basis.
- Associating bases with different vector models. Input is the vector model identifiers of the three bases; the comparison is whether the identifiers match; the observable result is that bases with differing identifiers must not be associated together, and only the exclusion and the actual observation are recorded.
How to verify the setup
- Input the three source sentences and the draft sentence. Compare the draft sentence "confirmed by the supplier" against the deduction notice value "pending supplier confirmation" and the reconciliation statement value "confirmation status not stated". The observable result is that the draft sentence is not supported by the source text and should be marked as pending verification. No live test is claimed here.
- Input the handover request "Please retrieve the three materials above and answer whether the deduction is confirmed". Compare whether the retrieved citations cover the three materials. When the citations contain all three materials, recall covers that scope; when any material is missing, record the gap without inferring a cause. For empty-result handling, a downstream design check can use the explicit paper input
[]; this does not assert that a fictional-rule retrieval necessarily returns[]. - Input the vector model identifiers of the three bases. Compare whether the identifiers match. Bases with matching identifiers may be associated together; bases with differing identifiers are excluded and recorded, and no cross-model retrieval is run to compare missed recall.
This page covers one scope only: fresh-food supplier claims, order fulfilment and after-sales question answering, knowledge base retrieval and recall. Whether materials, operations and requests in an operational handover are in an authorized state cannot be decided by retrieval or by a form; retrieval is not authorization.