What the documents look like
Fresh-food supplier claims pursue liability for quality at arrival. They stay separate from in-store natural shrinkage and markdown handling. The working materials come from quality inspection systems, WMS receiving records, supplier contracts and finance reconciliation. Day-to-day documents include quality exception forms, photo evidence, deduction notices and claim reconciliation statements. These materials update by batch and by month. The completeness of evidence for a single batch directly shapes later deduction calculation and negotiation over replacement or compensation.
Photo evidence and quality exception forms frequently arrive as a mix of images and text. Claim reconciliation statements carry fields such as batch number, receiving conclusion and deduction amount. Product reference and selection questions need these materials turned into searchable knowledge, so that procurement, quality inspection and supplier-facing staff can look up how earlier cases were handled.
Two features of this material shape the vector step. First, precise identifiers matter: batch numbers and deduction notice numbers must be findable, which favours full-text recall alongside semantic recall. Second, images are part of the evidence record, so whether an image can be retrieved depends on what was indexed at ingestion time, not only on what is uploaded at query time.
Constraints at the vector indexes step
Vector retrieval relies on an Embedding model to convert text and images into vectors, then computes distance to express similarity. Knowledge bases that are associated for retrieval must use the same vector model; check the actual model and the input types it accepts first. One original data block can be associated with multiple indexes, and the indexes and the original content are checked separately. Index content is built from complete and focused information, and retrieval coverage against index length is weighed against the sample. Whether an image can be retrieved depends on whether an image description or image vector was generated at ingestion; uploading an image at query time is not sufficient on its own.
Semantic recall and relevance scores do not carry permission or entity isolation. The same-vector-model requirement for associated retrieval libraries does not extend to business-registration sites or other external systems. The minimum relevance value is a filter threshold, not a final ranking rule. The citation limit is measured in tokens, which differs from a top k chunk count.
Hypothetical paper exercise sample: in a supplier claim knowledge base, the original text of one quality exception form reads: "On 12 March 2025, batch SP-0312 arrived with spinach showing 18% wilted leaf proportion in sampling; the receiving conclusion was partial rejection; deduction notice KC-0312 has been generated; the deduction amount is pending confirmation by finance reconciliation." A draft sentence awaiting judgment reads: "This batch has completed full deduction per contract and the case is closed." The materials needed for the check are the quality exception form above and the deduction notice, both within the accessible target retrieval scope. Compare same subject, same period, same field: the deduction notice shows the amount pending confirmation, while the draft sentence states full deduction completed and case closed. The two values are mutually exclusive on the same field, and the draft sentence is not supported.
Configuration and implementation checks
The table below centres on that sample and lists officially documented settings or implementation check items relevant to this feature and primary task. The values are hypothetical starting points for this exercise; the reason for choosing each and the comparison object are stated.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Vector model | Matches the model actually used by the knowledge base | When associating multiple knowledge bases, verify the identifier first and exclude those with a different model; do not run cross-model retrieval comparisons |
| Search mode | Hybrid search | Exact fields such as batch number and deduction number depend on full-text recall, while semantic questions depend on vector recall; the two are merged through RRF |
| Minimum relevance | 0.3 | Hypothetical starting point used to filter low-relevance results; compare against a raised recall count to observe whether irrelevant fragments decrease |
| Citation limit | 2000 tokens | Hypothetical starting point controlling cited volume given chunk length variation; compare against a lowered value to observe answer completeness |
| Result reranking | ReRank enabled | After hybrid search the candidate set is large, so the 0-1 rerank score is used for filtering; compare against the ordering when it is not enabled |
| Question optimization | Enabled for multi-turn follow-up scenarios | Completes references and expands semantics; compare against the retrieval effect in continuous dialogue when disabled, noting the extra model call latency |
Three common errors
- Treating semantic recall or a relevance score as permission or entity filtering. In the sample, deduction notice number KC-0312 belongs to a specific supplier. A retrieval hit on that content does not mean the current requester is entitled to view it. Authorization must be implemented downstream; retrieval itself does not constitute authorization.
- Treating image retrieval as a capability independent of the knowledge base. The sample contains photo evidence. If no image description or image vector was generated at ingestion, uploading an image at query time cannot produce a hit; check the ingestion index and model capability first.
- Writing RRF as a toggle, or treating an unstated field as confirmed. In the sample the deduction amount is pending finance reconciliation confirmation, while the draft sentence states full deduction completed and case closed. These are mutually exclusive values on the same field, so the draft sentence is judged unsupported. Do not fabricate authorization or effective clauses to reach a definite answer.
How to verify the setup
- Input the quality exception form original text above and the draft sentence, then compare the deduction field values: the original shows pending confirmation, the draft sentence shows full deduction completed and case closed. The two are mutually exclusive on the same field, so the draft sentence is unsupported and is recorded as pending verification.
- Input batch number SP-0312 and deduction number KC-0312, then compare the recalled content under hybrid search against pure semantic search: exact numbers are hit on the full-text recall path and may be missed on the semantic path. Observe whether the cited original text after RRF merging contains the corresponding numbers.
- Input a query containing an image only, then compare whether an image description or image vector was generated at ingestion: if not generated, a pure image query is essentially unusable; if generated, observe whether the image description path or the image vector path returns the corresponding photo evidence. An empty result can be checked against the downstream design using the explicit paper input [], without claiming that a fictional rule retrieval necessarily returns [].