What the documents look like
Fresh-food supplier claims pursue quality liability for goods as received. Store-level natural shrinkage sits outside this scope, so the material set is built around delivery and acceptance rather than shelf attrition. The data sources are quality inspection systems, WMS acceptance records, supplier contracts and finance reconciliation. The documents handled day to day are quality exception forms, photo evidence, deduction notices and claim reconciliation statements.
Two update rhythms coexist. Batch-level updates accompany each delivery, and monthly updates accompany reconciliation. A retrieval request may therefore land on a single batch of evidence or on a monthly reconciliation summary. The most time-consuming work in this category is fixing evidence, calculating deductions, and negotiating replacement or compensation. Knowledge-base retrieval and recall serves inventory procurement and supplier records: locating the original text of a batch quality exception form, the descriptive text of the corresponding photo evidence, the amount basis in a deduction notice, and the period scope of a claim reconciliation statement. A retrieval result does not determine liability and does not substitute for authorization.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. Verify the identifier first and exclude those that do not match; do not run cross-model retrieval to compare missed recalls. Retrieval searches the whole base. A collection is used only for classification and management, and has no bearing on search results.
A single retrieval combines question optimization, semantic recall, full-text recall, image retrieval and reranking according to the input content and search parameters, then fuses the multiple result paths through RRF. Rerank requires an available model and must be enabled by configuration. Image retrieval requires corresponding image input and model capability. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Cited content is output as an array whose length may be zero; when nothing is found, the output path still completes.
Semantic recall and relevance scores do not carry permission or entity isolation. The same-vector-model requirement for associated bases does not extend to business registration sites or other external systems. Retrieval is an action that follows authorization. Forms collect values; business verification, approval, database writes and external synchronization each require their own stated downstream implementation.
Hypothetical exercise sample. Material text for comparison: a supplier contract annex states that "a quality exception form must be submitted within 48 hours after arrival acceptance; if overdue, the claim is deemed waived." A batch quality exception form records "arrival date 12 March, acceptance date 14 March, exception described as wilted leafy greens, submission date 18 March." Draft sentence to be assessed: "The claim for this batch was submitted late and is deemed waived." This sentence cites only these two materials; no deduction notice or claim reconciliation statement is attached.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge base | Inventory procurement and supplier records base, single base | Verify the vector model identifier first; if a candidate base carries a different identifier, exclude it and do not associate them together |
| Search mode | Hybrid retrieval | Compare hit results for the same test question against pure semantic and pure full-text search |
| Citation limit | Hypothetical starting point of 2000 tokens | Test with quality exception forms and reconciliation statements of differing length to see whether content stays complete, then compare against 1000 and 4000 |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then compare the filtered sample distribution against 0.3 and 0.5 |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enabled | Confirm the rerank model is available; compare candidate order and scores before and after reranking |
These values are hypothetical starting points for this exercise. They are not official defaults and not limits specified in the documentation.
Three common errors
- Treating a retrieval hit as an established claim. In the sample, the quality exception form shows a submission date later than the 48-hour range in the contract annex. Retrieval can recall both original texts, but the deduction notice and claim reconciliation statement do not appear. The draft sentence "late and is deemed waived" is supported only by those two materials; the consequence of lateness is not stated in the sample and remains unsupported.
- Associating bases across different vector models. When two bases carry different vector model identifiers, they are not associated together. Verify the identifiers, exclude the non-matching base and record what was observed; do not infer missed recalls or redundancy.
- Treating a pending request as executed. A handover request in pending-review status indicates only its position in a process. It does not demonstrate that operating permission was obtained, and it does not demonstrate that a termination was completed or that compensation was paid.
How to verify the setup
- Input the contract annex text and the quality exception form text from the hypothetical sample. Compare whether the subject, batch and date fields in the two materials point to the same object, and observe whether retrieval recalls both and returns non-empty cited content. If only one is recalled, record the gap and do not fabricate a clause.
- Input the draft sentence "The claim for this batch was submitted late and is deemed waived." Compare the scope of that sentence against the wording of the two original texts, and observe whether it asserts a consequence not stated in the originals. If the originals do not state the consequence of lateness, judge the sentence unsupported and preserve the evidence gap.
- Input an explicit paper-based empty-result scenario and check whether the downstream design accepts a cited array of length zero, observing whether the output path still completes. Do not assert that a fictional rule search must return an empty array, and do not guarantee a first hit or complete recall.