What the documents look like
Fresh-food supplier claims pursue responsibility for quality at arrival and do not handle natural in-store shrinkage. That distinction shapes which materials retrieval has to reach: quality anomaly sheets, photo evidence, deduction notices and claim reconciliation statements. Cross-border suppliers and multilingual text are common, so inspection conclusions, receiving records and contract clauses sit in different languages, and retrieval has to locate the responsibility basis for the same batch across those languages. Materials update on two rhythms: quality anomaly sheets are produced per batch, while claim reconciliation statements are gathered monthly. Because the two rhythms differ, retrieval scope and citation limits need separate checks.
A second feature matters for search mode choice. Quality anomaly sheets carry natural-language descriptions of what was found, while batch numbers, deduction notice numbers and supplier codes appear as exact strings. Semantic retrieval suits the first kind of content through vector distance; full-text retrieval suits the second through keyword hits. Mixed retrieval combines both paths and merges them with RRF, after which ReRank can reorder candidates. A third feature matters for filtering. The minimum relevance value is a threshold that applies only under semantic retrieval or when result re-ranking is used; it is a similarity filter and does not act as a permission or entity isolation control. Retrieval returns citation content only. Deduction execution, claim confirmation and supplier notification are downstream business actions that need their own implementation.
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, so the vector model identifier has to be checked first and non-matching bases excluded. Retrieval returns citations as an array whose length can be zero; the downstream chain still runs when nothing is found. Semantic retrieval supports cross-language understanding, which helps when a Chinese question has to match an English knowledge point. Full-text retrieval fits batch numbers, deduction notice numbers and supplier codes. Mixed retrieval merges multiple recall paths through RRF and then applies ReRank. The minimum relevance value takes effect only under semantic retrieval or result re-ranking, and it filters low-relevance results rather than isolating entities or granting access.
Hypothetical paper exercise sample. A quality anomaly sheet for one batch reads: "Batch A-0312 arrived on 12 March 2024. Sampling found a wilting ratio in leafy vegetables above the acceptance standard. Responsibility is determined to rest with the supplier. A deduction of CNY 8,000 is proposed, pending review by the quality inspection supervisor." Photo evidence for the same batch states: "Taken on 12 March 2024 at the unloading area, showing wilting in leafy vegetables." A drafted sentence reads: "This batch has completed the deduction and the supplier has been notified." The question is whether that drafted sentence is supported by the materials. The source text records a proposed deduction and a pending supervisor review; it does not record a completed deduction or a supplier notification.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the quality claim base that uses the same vector model | Bases on different vector models cannot be associated together; check the vector model identifier and exclude non-matching bases |
| Search mode | Mixed retrieval | Quality anomaly sheets contain natural-language description and batch numbers, so both the semantic and full-text recall paths need checking |
| Citation limit | Start at 3,000 tokens | Reconciliation statements have long passages; test with materials of clearly different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Start at 0.5 | Label relevant and irrelevant samples first, then compare whether filtering drops key passages from the citations |
| Question optimization | Enabled | On follow-up questions such as the basis for this batch deduction, check whether the subject is altered |
| Result re-ranking | Enabled, with re-rank score checked | When candidates are numerous, compare ordering before and after re-ranking to see whether the quality anomaly sheet moves forward |
These values are hypothetical starting points for this exercise and are not official defaults. The citation limit uses tokens as its unit, not an item count. Question optimization runs a model call before retrieval, which adds latency; whether to keep it on depends on observed results for the actual question pattern.
Three common errors
- Treating the citation limit as a top-k item count. The limit is measured in tokens, and chunk lengths vary widely, so an item count produces unstable results. Check: input two reconciliation passages of clearly different length, observe whether the cited content stays complete, and compare the token measure against an item count as two separate readings.
- Assuming the minimum relevance value excludes records belonging to unrelated suppliers. It is a filter threshold that applies only under semantic retrieval or result re-ranking and does not isolate entities. Check: run a query containing an unrelated supplier code, observe whether that code appears in the returned citations, and record what is observed rather than inferring permission behavior.
- Treating retrieval output as deduction authorization. Retrieval returns citation content only; deduction execution, claim confirmation and notification are downstream actions. Check: compare the drafted sentence "This batch has completed the deduction and the supplier has been notified" against the source text "pending review by the quality inspection supervisor." The same subject, the same period and the same status field carry mutually exclusive values, so the drafted sentence is unsupported.
How to verify the setup
- Input the hypothetical quality anomaly sheet text and the drafted sentence given above, then compare the status field in each. The source records "pending review by the quality inspection supervisor," while the drafted sentence states that the deduction is complete and the supplier notified. Same subject, same period, same field, mutually exclusive values. When that condition holds, the drafted sentence is recorded as unsupported.
- Input the batch number "A-0312" using full-text retrieval and compare whether the returned citations include the quality anomaly sheet for that batch. When the condition holds, the citation content contains that sheet and the citation array length is greater than zero. If an empty array is returned, record it as not recalled and do not infer a cause.
- Input the cross-language question "What is the deduction basis for batch A-0312?" and compare the citations returned by semantic retrieval and by mixed retrieval. When the condition holds, the citations contain the determination statement and the deduction amount from the quality anomaly sheet and do not contain other supplier codes. If a key passage is missing, record the gap and do not assert that retrieval must fail.
An empty result can be checked against a clearly labelled paper input of an empty array to confirm how the downstream design handles it; this does not show that any particular query returns an empty array. Whether image retrieval or re-ranking runs depends on the actual input and on model capability and settings. For this page, the primary task is retrieval and recall, and each check above compares a named input against a named source field or returned citation.