What the documents look like
Fruit-store procurement runs on wholesale-market quotations and grade-based acceptance, which differs from head-office standard-meal ordering in convenience retail. The documents involved are purchase orders, acceptance records, pesticide-residue test reports and loss records, produced daily or per batch. Retrieval therefore works across the same supplier and the same batch appearing in several documents, and the subject, period and field must line up before any judgment is possible. Document length varies widely, from a one-line quotation to a full page of test conclusions, so the reference cap is expressed in tokens rather than item count.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; the model identifier is checked first and non-matching bases are excluded. Semantic retrieval, full-text retrieval and hybrid retrieval are compared against test questions, and no single mode is assumed to suit all material. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion. Reference content is output as an array, and its length may be zero. RRF is the fusion step for hybrid results. Image retrieval requires a corresponding image input and model capability, and reranking requires a rerank model enabled under the settings. Retrieval does not perform authorization judgment.
Hypothetical exercise sample: a store's fresh-produce supervisor keeps a paper record whose source text reads: "Supplier A, quotation dated 10 March 2025: Red Fuji apples, 80 mm and above, 10 kg per box, quotation valid until 17 March 2025. The acceptance record of the same day notes 30 boxes of this batch arrived and the grade passed." The draft sentence to be assessed is: "Supplier A has confirmed supply at this price for the whole of March 2025." That sentence treats the quotation validity period as a supply period and states no whole-month supply commitment.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only bases containing purchase orders, acceptance records, test reports and loss records, with matching vector-model identifiers | Check the vector-model identifier first and exclude non-matching bases |
| Search mode | Start with hybrid retrieval | Compare hit situations against semantic retrieval alone and full-text retrieval alone |
| Reference cap | Set to a token amount covering one complete acceptance record | Compare against a shorter cap and observe whether reference content is truncated |
| Minimum relevance | Mark relevant and irrelevant samples first, then set the threshold | Compare whether irrelevant fragments are filtered out at different thresholds |
| Question optimization | Enable for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Use when a rerank model is available and enabled under the settings | Compare candidate ordering before and after enabling, and observe score-based filtering |
These values are hypothetical starting points for this exercise, not official default values.
Three common errors
- Treating the quotation validity period as a supply period. In the sample, "quotation valid until 17 March 2025" states only the range in which the quotation holds, while the draft sentence writes "supply at this price for the whole of March," which is an inference beyond the source text. The period field in the draft sentence should be compared with the source text to confirm whether a whole-month supply commitment is stated.
- Treating a retrieval hit as authorization confirmation. Retrieval returning material about Supplier A does not mean that supplier has received procurement authorization or that a contract has taken effect. Retrieval results must be distinguished from downstream actions such as signing and approval.
- Treating an unstated item as a negation. The sample does not state a whole-month supply commitment, so this can only be recorded as unsupported or pending verification, not as a conflict or a negation. The same subject, the same period and the same field should be checked for mutually exclusive values.
How to verify the setup
- Input the sample source text and the draft sentence, and compare the period wording of the draft sentence with "quotation valid until 17 March 2025" in the source text. If the draft sentence contains whole-month supply wording while the source text does not state it, the result is beyond the source text and the draft must be returned for revision.
- Input the sample source text and search for "Supplier A acceptance grade." Compare whether the returned reference contains the acceptance record's subject and batch. If the reference content is an empty array, handle the downstream design according to an empty result; do not infer that the rule will necessarily return empty.
- Input the identifiers of two knowledge bases built on different vector models and compare whether the identifiers match. If they do not match, exclude the non-matching base and record the actual observation; do not run a cross-model retrieval to compare missed recalls.