What the documents look like
Cold-chain receiving records are produced around transport temperature, arrival batches and store receiving actions. Temperature data from refrigerated vehicles, WMS entries, store receiving system records, quality inspection photographs, plus cold-chain temperature logs, arrival acceptance sheets, rejection sheets and supplier delivery notes together form the retrieval scope. Materials update per batch and per day, and the same supplier arriving on different dates produces separate records for each batch.
After these materials enter the knowledge base, the smallest search unit is the library; collections serve classification management only. One data group can map to multiple vectors, and a hit on any one vector recalls that data group. In store service and internal training, retrieval helps receiving staff locate the temperature record, acceptance conclusion and rejection basis for the corresponding batch.
Constraints at the retrieval step
A single batch involves a temperature record, an acceptance sheet and possibly a rejection sheet. A search term such as "acceptance" alone can recall same-type documents from other batches. Documents contain precise fields such as batch number, supplier name and arrival date. Full-text retrieval suits number-type queries, while semantic retrieval suits natural-language descriptions. Hybrid retrieval merges multiple result paths through RRF. ReRank requires a model and must be enabled by configuration; its score is 0-1 and can be used for filtering. The citation upper limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or result re-ranking. The dataset search node can associate one or more knowledge bases that use the same vector model; libraries with different vector models cannot be associated together. Retrieval results do not perform authorization checks and do not isolate unauthorized records.
Hypothetical drill sample: the source text for comparison reads: "On 12 March 2025, supplier A delivered batch A-0312 to the store. The refrigerated vehicle temperature record shows 2°C to 5°C throughout. The store receiver completed sampling that day, and the arrival acceptance sheet registered the batch as accepted into stock." A draft sentence to be assessed reads: "Supplier A batch A-0312 was rejected because the temperature exceeded the limit." The draft sentence and the source text differ in the acceptance conclusion field: the source registers accepted into stock, while the draft states rejection. Same subject, same batch, same date, and the two values conflict.
Configuration and implementation checks
| Configuration item | Hypothetical starting point and reason for this drill | Verification method |
|---|---|---|
| Associated knowledge bases | Associate only the knowledge base holding this batch's temperature record and acceptance sheet | Check whether the vector model identifiers of the libraries match; exclude those that do not |
| Search mode | Start with hybrid retrieval to compare number-type and natural-language questions | Search separately with the batch number and a descriptive question; observe recalled content |
| Citation upper limit | Hypothetical starting point sized to cover one acceptance sheet and one temperature record | Compare against actual materials of differing length; observe whether content is complete |
| Minimum relevance | Hypothetical starting point at a middle value, to observe filtering behaviour | Label relevant and irrelevant samples first, then compare different values |
| Question optimization | Enable for multi-turn follow-ups containing references; disable for explicit questions | Compare the original question with the optimized question; check whether the subject was altered |
| Result re-ranking | Enable when a re-ranking model is available and candidates are numerous | Observe whether the re-ranking score is used for filtering; record the actual ordering |
Three common errors
- Treating a retrieval hit as authorization. Retrieval returns citation content; business approval, database writes and external synchronization each require their own downstream implementation. A retrieval result does not constitute authorization to release or reject a batch.
- Associating knowledge bases across vector models. The node requires knowledge bases with the same vector model. When model identifiers differ, exclude the mismatched library, record the actual observation, and do not run cross-model retrieval to compare missed recalls.
- Treating a draft sentence as a verified conclusion. In the sample, the draft sentence states the batch was rejected, while the source registers accepted into stock. The two differ in the same field, which is a conflict pending verification, and rejection cannot be treated as having occurred on that basis.
How to verify the setup
- Input the hypothetical source text and the draft sentence, then compare the acceptance conclusion field. The source reads accepted into stock; the draft reads rejected. Under the same subject, same batch and same date, the values differ, so this is judged a field conflict and requires returning to the original acceptance sheet for verification.
- Input batch number A-0312 and a descriptive question, then compare whether the recalled content points to that batch. If the recalled content contains other batches or lacks the temperature record, record it as not hitting the target batch, adjust the search mode and observe again.
- Input a question with no corresponding material and observe the citation content array. The documentation states the citation length can be 0 and the output path still proceeds even when nothing is found. Use a clearly stated paper input to check the downstream design; do not assert that a fabricated rule necessarily returns an empty array.