What the documents look like
Fruit-store procurement records center on purchase orders, acceptance sheets, pesticide-residue test reports and loss records. Their sources include wholesale-market price information, supplier quotations, store sales and test reports, and they are refreshed daily or per batch. In cross-border and multilingual retrieval, the same batch may carry acceptance grades, residue conclusions and loss entries written in different languages, so a search has to reach the matching batch across those languages. Semantic retrieval can match a Chinese question to an English knowledge entry, which connects directly to this mixed-language reality. Because records change frequently, a hit must correspond to the current batch. The reference output is an array whose length may be zero, and the output path still runs when nothing is found.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. The smallest search unit is the base; a collection only classifies data and does not affect search results. Semantic retrieval relies on vector distance and suits natural-language questions and semantically close content. Full-text retrieval relies on keyword hits and suits exact queries such as numbers, models, proper nouns and error codes. Hybrid retrieval uses both paths and merges them through RRF. The minimum relevance value is a filter threshold that takes effect only in semantic retrieval or when result reranking is used; it does not isolate permissions or entities, and retrieval itself is not authorization. The reference cap is measured in tokens and limits how much content one search may cite. Question optimization adds one model call before retrieval for coreference removal and question expansion, which increases overall latency when enabled. Image retrieval needs image input and matching model capability, and reranking needs an available model enabled by configuration; these optional paths do not mean every request runs all of them.
Hypothetical paper test case: a store's fresh-produce supervisor submits an English acceptance sheet scan reading "Batch No. GZ-2024-0517, Grade A navel oranges, 320 cartons, received 2024-05-17, cold chain intact." A Chinese pesticide-residue test report reads "Sample No. GZ-2024-0517, test date 17 May 2024, chlorpyrifos not detected." The draft sentence to be judged is: "Batch GZ-2024-0517 has passed acceptance and pesticide-residue testing and may be put on sale." The draft merges an acceptance fact and a test conclusion into a sale authorization. Neither the acceptance sheet nor the test report states any listing authorization, so that conclusion is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the fruit-procurement base with the same vector model | Check each base's vector-model identifier, exclude bases whose identifiers differ, and do not run cross-model retrieval |
| Search mode | First test semantic retrieval and full-text retrieval separately | Compare the recalled content of both modes against the same GZ-2024-0517 materials |
| Reference cap | Set 2000 tokens as the hypothetical starting point for this trial | Test with acceptance sheets and test reports of different lengths, and compare completeness against a shorter cap |
| Minimum relevance | Set 0.5 as the hypothetical starting point for this trial | First label samples related and unrelated to GZ-2024-0517, then observe how the value affects results |
| Question optimization | Leave it off in this hypothetical trial | Compare the original question with an optimized question, check whether the subject was altered, then decide whether to enable it |
| Result reranking | Not enabled in this hypothetical trial | Confirm the rerank model is available and enabled by configuration before comparing ranking changes |
Three common errors
- Treating a retrieval hit as authorization to act. The input is the draft sentence "Batch GZ-2024-0517 has passed acceptance and pesticide-residue testing and may be put on sale." The comparison objects are the acceptance sheet text and the test report text, and neither states listing authorization. Retrieving these two materials produces no authorization effect, and listing is a business action requiring separate system implementation.
- Associating knowledge bases that use different vector models. The input is two bases whose vector-model identifiers differ. The comparison object is each base's vector-model identifier; different identifiers mean exclusion, and no cross-model retrieval is run to compare missed recalls.
- Treating minimum relevance as a permission or entity filter. The input is a minimum relevance of 0.5. The comparison object is the score distribution of related and unrelated samples. This value only filters low-relevance results and does not exclude unrelated businesses or unauthorized records.
How to verify the setup
- Input the acceptance sheet text "Batch No. GZ-2024-0517, Grade A navel oranges, 320 cartons, received 2024-05-17, cold chain intact." and the Chinese question "acceptance grade for batch GZ-2024-0517." Compare whether the returned reference content contains Grade A and 320 cartons. The condition is met when the reference content contains that batch's acceptance information. If an empty array is returned, record that result and do not infer a cause.
- Input the draft sentence "Batch GZ-2024-0517 has passed acceptance and pesticide-residue testing and may be put on sale." Compare the scope of the draft sentence with the acceptance sheet and test report text. The condition is met when neither source states listing authorization, and the draft conclusion remains unsupported.
- Input the Chinese question "chlorpyrifos test result" and the English test report. Compare what semantic retrieval and full-text retrieval return. The condition is met when semantic retrieval reaches the cross-language matching content. If full-text retrieval does not match, record the difference between the two modes without guaranteeing a first hit or full recall.