What the documents look like
Cold-chain receiving at a store turns on transport temperature, arrival batch, and quality sampling. It does not cover origin price negotiation, which belongs to fruit purchasing. The working sources are cold-chain vehicle temperature control, WMS, the store receiving system, and quality inspection photos. The documents handled each day are cold-chain temperature records, arrival acceptance forms, rejection forms, and supplier delivery notes. Records update per batch and per day.
Two features of these materials shape retrieval. First, a single batch carries several linked fields: the temperature value, the acceptance conclusion, and the rejection reason. A retrieval result that mixes the conclusion of one batch with the reason of another is not usable at the receiving desk. Second, the questions asked of the knowledge base fall into two kinds. One kind is precise: a batch number, a plate number, a temperature range. The other kind is natural language, such as whether a delivery can be accepted. The documented search modes treat these differently, so the retrieval setup has to be checked against both.
The public brief lists the three most time-consuming steps as measuring temperature, sampling quality, and registering the rejection reason. Those steps are the operational context for the records, not functions of the retrieval node itself.
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 check at this step is to confirm the vector model identifier and exclude those that do not match. A cross-model search is not run in order to compare what was missed.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Batch numbers, model numbers, and proper nouns suit full-text or hybrid retrieval. Natural-language questions suit semantic or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion step, not a standalone switch. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it acts as a filter threshold. It does not perform permission filtering or entity isolation. Retrieved references are output as an array whose length can be zero, and the downstream link still runs when nothing is found.
Image retrieval requires the corresponding image input and model capability. Reranking requires a rerank model and that it be enabled by configuration. The optional paths drawn in the overall retrieval flow are not all executed on every request.
A hypothetical paper exercise sample is used below. The stored text of an arrival acceptance form reads: "Batch LC2024-0312, the cold-chain vehicle temperature record shows an arrival temperature of 2 degrees Celsius, three crates of leafy greens were sampled, two crates passed, one crate showed wilted leaves, the acceptance conclusion is partial rejection, the rejection reason is wilted leaves." The draft sentence to be checked reads: "This batch was rejected in full and the supplier raised no dispute." The draft and the source give different values for the same field, on the same subject and the same batch.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Basis and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the cold-chain receiving library that uses the same vector model | Check the vector model identifier first and exclude libraries on a different model |
| Search mode | Hybrid retrieval | Observe recall separately with a batch-number question and a natural-language question |
| Reference limit | Start at a token amount that covers the full text of one acceptance form | This is a hypothetical starting point for this exercise; compare against shorter and longer documents for completeness |
| Minimum relevance | Start at 0.3 | This is a hypothetical starting point for this exercise; compare against labelled relevant and irrelevant samples |
| Question optimization | Enabled for multi-turn follow-up questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled when a rerank model is available | Observe the rerank score and any change in candidate ordering |
Three common errors
- Associating knowledge bases that use different vector models. The input is two knowledge bases with different vector model identifiers. The comparison is whether the identifiers match. The observable result is that the non-matching library is excluded and no cross-model search is run.
- Treating the minimum relevance value as a permission filter. The input is material for one batch plus an unrelated query. The comparison is the threshold setting against the returned content. The observable result is that the threshold affects relevance filtering only and does not exclude records that were not authorized.
- Treating a rejection form awaiting review as a completed rejection. The input is the source text of a rejection form marked pending review. The comparison is the status field against the draft conclusion. The observable result is that the status remains pending review and the completed rejection in the draft is not supported.
How to verify the setup
- Enter the acceptance form source text and the draft sentence from the hypothetical sample above. Compare the value of the acceptance conclusion field. The draft and the source give different values, so the draft is judged inconsistent with the source and must be corrected to match it.
- Enter the batch number "LC2024-0312" and the natural-language question "can this batch of leafy greens be accepted". Compare whether the recalled content under both inputs contains the acceptance form for that batch. If both inputs recall the same batch material, hybrid retrieval is usable for this sample. This is a paper check and does not assert that the first result always matches or that every relevant item is recalled.
- Enter a clearly specified paper input of an empty result array []. Compare whether the downstream reference handling still runs. The observable result is that the empty-result link executes. This does not assert that retrieval under any invented rule always returns [].
All three steps above are paper checks. No deployment or execution took place, and no approval, refund, or other business action is claimed.