What the documents look like
Selection questions in fresh-fruit store procurement draw on wholesale market quotations, supplier quotes, store sales, and pesticide-residue test reports. Purchase orders, acceptance sheets, pesticide-residue test reports, and loss records form the main document set. These materials determine whether a selection question can be answered and whether the cited source text in an answer can be checked. The materials are produced per batch, so the same product may carry different quotes and test conclusions across batches. Retrieval and citation must land on a specific batch and date range.
The knowledge-base search node can associate one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. Once citation content is passed to the answering step, the answer and the cited source text must be checked line by line. The minimum relevance value 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. RRF is a fusion process for mixed-retrieval results, not a toggle. Image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled by configuration. Optional paths shown in the document flow diagram do not mean every request executes all of them.
Constraints at the citation tracing step
A worked paper sample is given here for checking. The source text to be checked reads: "On 10 April 2025, a store asked stall A for a quote on Hongyan strawberries; the quote sheet states 12 yuan per jin, valid until 12 April 2025; the pesticide-residue test report for that day is numbered A-0410 with a conclusion of qualified." The draft sentence to be judged reads: "Stall A on 10 April confirmed supply at 12 yuan through the end of April." This sentence disagrees with the source on the supply-period field. The source states only a quote validity period and does not state a supply period, so the draft sentence is unsupported.
Three points follow from this. A quote validity period is not a supply period. Associating knowledge bases that use different vector models cannot be followed by inferring missed recall, redundancy, or unpredictable behavior; only the identifiers are checked and non-matching ones are excluded, with actual observations recorded. A relevance score or semantic score is not a permission decision; an unauthorized record is not excluded by a similarity threshold, and permission requires a separate system implementation.
Configuration and implementation checks
| Configuration item | Setting or check content | Relation to this page's documents |
|---|---|---|
| Associated knowledge bases | Check the selected knowledge-base identifiers and whether the vector model is the same; exclude bases with a different vector model | Cross-model bases cannot be associated together |
| Search mode | Choose semantic retrieval, full-text retrieval, or hybrid retrieval according to question type | Number and model questions suit full-text or hybrid retrieval |
| Citation upper limit | Set the maximum content cited per search, measured in tokens | For this hypothetical trial the starting value may be set close to the length of one test report, then compared with the default before deciding |
| Minimum relevance | Acts as a filter threshold only under semantic retrieval or result reranking | Not used to filter unrelated businesses or unauthorized records |
| Question optimization | Decide whether to enable it for follow-up turns and short questions | Enabling it adds one model call and additional time |
| Citation output | Accept an array of length zero and design the downstream handling | An empty array does not mean the workflow has errored out |
Three common errors
- Treating a quote validity period as a supply period. The input is the quote sheet above and the draft sentence; the comparison is between the field "valid until 12 April" and the field "supply through the end of April"; the observable result is that these are different fields with different values, so the draft sentence is unsupported.
- Associating knowledge bases with different vector models and then directly inferring missed recall or redundancy. The input is two knowledge-base identifiers; the comparison is the vector-model identifier; the observable result is that different identifiers exclude the association, while actual retrieval behavior is left to be recorded without a preset conclusion.
- Using the minimum relevance or a semantic score in place of a permission decision. The input is one unauthorized record and one low-score record; the comparison is the filtering basis; the observable result is that relevance affects only ordering and the filter threshold, does not exclude the unauthorized record, and permission needs a separate system implementation.
How to verify the setup
- Input the quote sheet source text and the draft sentence from the hypothetical sample; compare the "validity period" field and the "supply period" field; observe whether the draft sentence is marked unsupported. If it is marked confirmed, the check fails.
- Input two knowledge-base identifiers that use different vector models; compare whether the model identifiers match; observe whether the system prevents joint association. If association is allowed and missed recall is asserted from it, the check fails.
- Input the paper value
[]as the citation output; compare whether the downstream node proceeds; observe whether the empty result is treated as an error that terminates the flow. If the downstream stops because of the empty array, the check fails.
These checks state expected conditions and do not claim that any test has been run. The hypothetical sample keeps its evidence gaps: the source does not state a supply period, and that gap remains open rather than being filled by an assumed authorization or effective clause.