What the documents look like
Cold-chain receiving at a store produces records on a per-batch and per-day rhythm. The working materials are cold-chain vehicle temperature records, arrival acceptance sheets, rejection notes and supplier delivery notes. The receiving step involves measuring temperature, sampling quality and registering the reason for any rejection. Errors in this category can lead to spoiled ingredients, supplier disputes and regulatory penalties. This page addresses knowledge-base retrieval for operating reports and anomaly reviews, where staff need to pull temperature ranges, batch numbers, rejection reasons and acceptance conclusions out of those materials as cited excerpts for review.
Two features of these materials shape the retrieval task. Temperature records carry numbers and units, while batch numbers and rejection-note numbers are exact-match objects; these suit full-text or hybrid retrieval. Acceptance conclusions and anomaly descriptions are natural language and suit semantic retrieval. A single retrieval configuration does not fit all of this material, so the mode is chosen against the actual document shapes rather than a uniform setting.
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; the identifiers must be checked first and non-conforming bases excluded. Cited content is output as an array whose length can be 0, so the output path still runs even when nothing is found.
The minimum relevance score is a filter threshold that takes effect only under semantic retrieval or when result reranking is used. It does not perform permission or entity isolation, and it is not a final ranking rule. Image retrieval requires image input together with the corresponding model capability, and reranking requires a ReRank model enabled according to the configuration. The paths shown in the overall retrieval flow do not all execute on every request.
Hypothetical paper test case. A store receives a batch of chilled pork on 12 March 2025. The temperature record states: "Compartment temperature 4.2°C, arrival time 06:40, batch A-0312." The arrival acceptance sheet states: "Sampled 3 boxes, core temperature 5.1°C, judged acceptable." The rejection note field is empty. A draft sentence awaiting judgment reads: "This batch was rejected because the temperature exceeded the limit." Judgment requires the same subject, the same period and the same field. Neither the temperature record nor the acceptance sheet contains a rejection entry, and the rejection note is empty, so the draft sentence is unsupported and remains pending verification.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to this page's sample. The numeric values are hypothetical starting points for this exercise, with the reason for choosing them and the comparison target stated.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Cold-chain temperature record base, arrival acceptance sheet base | Both bases must use the same vector model; check identifiers first and exclude non-conforming ones |
| Search mode | Hybrid retrieval | Batch numbers go through full-text, anomaly descriptions through semantic; compare the three modes against test questions |
| Citation limit | 1500 tokens as the starting point for this exercise | The unit is tokens; test content completeness with materials of different lengths, then adjust |
| Minimum relevance score | 0.5 as the starting point for this exercise | Takes effect only under semantic retrieval or result reranking; label relevant and irrelevant samples first, then tune |
| Question optimization | Enabled | Used for coreference resolution and question expansion; compare the original and optimized question to check whether the subject was altered |
| Result reranking | Enabled | Requires a ReRank model; the rerank score is 0-1 and filtering can use that score |
Three common errors
- Treating the minimum relevance score as a permission or entity-isolation mechanism. Input is a cross-store query request; compare the retrieval configuration against the authorization boundary. Expecting a threshold adjustment alone to exclude unauthorized records is an error, because retrieval does not carry authorization.
- Associating knowledge bases built on different vector models. Input is the identifiers of two knowledge bases with different vector models; compare the identifiers against the model field. Associating them without checking the identifiers is an error; non-conforming bases should be excluded and the actual observation recorded.
- Reading an empty rejection note as a completed rejection. Input is the empty rejection note field in the sample; compare the acceptance sheet and rejection note fields for the same subject and period. Writing the draft sentence as "rejected because the temperature exceeded the limit" is unsupported; missing evidence stays missing.
How to verify the setup
- Input the sample source text and the draft sentence; compare the same-field values across the temperature record, the acceptance sheet and the rejection note, and observe whether the draft sentence contradicts any material. When none of the three materials contains a rejection entry, the judgment is unsupported and is not written as a negation.
- Input the batch number "A-0312" and the natural-language question "how did the acceptance of this batch of meat go"; compare the content recalled by the full-text and semantic paths, and observe whether the citation array contains the corresponding batch excerpt. If one path is empty, record the empty result and check the downstream design; do not claim that a rule-based retrieval necessarily returns an empty array.
- Input a query containing image input; compare whether the knowledge base has a vision model configured or a vector model that supports images, and observe whether an image index was built at ingestion. If it was not built, record the gap; do not infer missed recalls or redundancy as runtime results.
This page addresses only store cold-chain receiving, operating reports and anomaly reviews, and knowledge-base retrieval. At operational handover, which materials, operations and requests sit at which authorization state cannot be treated as retrieval or a form performing authorization. The retrieval check describes observations and failure branches; it does not guarantee a first-hit result, full recall, or that an unrelated query returns empty. An empty result can be checked against downstream design using a clearly stated paper input of [], without claiming that a hypothetical rule-based retrieval necessarily returns []. Knowledge bases on different vector models cannot be associated together; check the identifiers, exclude non-conforming bases and record the actual observation, without executing a cross-model retrieval to compare missed recalls. Optional paths shown in the overall document flow do not all execute on every request: image retrieval requires the corresponding image input and model capability, and reranking requires the model to be present and enabled.