What the documents look like
Complaints and official replies are directed at a platform or a regulator, and they call for formal evidence and corrective wording. The working material arrives from the 12315 platform, the platform complaint centre, customer service chat and the order system. The documents handled day to day are complaint tickets, regulator inquiry letters, order evidence and rectification reports.
For this page the retrieval target sits inside inventory procurement and supplier records. Once those records enter a knowledge base, search is aimed at procurement batches, supplier entities, supplied quantities, arrival dates and authorization status. Material is updated irregularly, case by case. When the records of a case lack a shared numbering system, the quality of the search terms directly affects what is recalled.
Two features of this material shape the retrieval step. First, the same case may combine a complaint ticket, an inquiry letter and an order record, so a single question can need evidence from more than one document. Second, exact strings such as a purchase order number or a batch identifier behave differently from a natural-language description of what happened, and the two need different search modes.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together; the model identifier of each base is checked and any base that does not match is excluded. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Identifiers, model numbers and proper nouns suit full-text or hybrid retrieval, while natural-language questions suit semantic or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and a rerank model reorders candidate text, producing a score from 0 to 1. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The citation limit is measured in tokens and controls the length of returned content. Question optimization performs coreference removal and question expansion and adds one model call. Image retrieval needs image input and matching model capability, and reranking needs a model and is enabled according to settings. The output of the node is an array of citations whose length may be 0, which means the output path continues even when nothing is found.
Hypothetical drill sample: a shop receives a regulator inquiry letter asking it to explain the source of a batch of children's toys. The material text to be checked reads: "On 10 March 2025, this shop purchased 200 children's building blocks from supplier A, purchase order CG-20250310-07. A copy of supplier A's business licence and the factory inspection report for this batch are attached." The draft sentence to be judged reads: "This batch of goods has obtained supplier A's authorization, and the supply validity period runs to 31 December 2025." The source text records only the purchase fact and the attached certificates; it states no authorization clause and no supply validity period.
Configuration and implementation checks
| Configuration item | Setting for this page | How to check |
|---|---|---|
| Linked knowledge bases | Select bases holding complaint tickets, regulator inquiry letters, order evidence and rectification reports that share the same vector model | Compare the vector model identifier of each base and exclude any that do not match |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | Use a purchase order number question for full-text, a descriptive question for semantic |
| Citation limit | For this hypothetical trial the starting value is 2000 tokens, because the sample contains several evidence passages | Compare against shorter material and observe whether content is truncated |
| Minimum relevance | For this hypothetical trial the starting value is 0.3, after labelling relevant and irrelevant samples | Compare the relevance of returned items at different thresholds |
| Question optimization | Used for coreference removal and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled according to settings when a rerank model is available | Compare candidate ordering before and after enabling |
The two starting values above are hypothetical trial values chosen for this sample. They are not documented defaults, and the comparison method for each is given in the third column.
Three common errors
- Treating minimum relevance as a permission filter. In the hypothetical sample, purchase order CG-20250310-07 belongs to supplier A. A search returning that order number shows semantic relevance only; it does not show that the requesting party is entitled to view that supplier's records. Check: enter the order number as a query, compare the returned content against the access permission list, and observe whether unauthorized records appear.
- Treating a retrieval hit as confirmed authorization. The draft sentence states that supplier A's authorization has been obtained, while the source text records only the attached business licence and inspection report. Check: enter the draft sentence together with the source text, compare whether the word "authorization" has a corresponding field, and observe whether the source states an authorizing entity and scope.
- Treating a pending state as already effective. Suppose a handover request reads "authorization material has been submitted, an external reply may be sent." Submission means only that material entered a review process. Check: enter that request together with the material status field, compare whether "submitted" and "approved" are the same value, and observe whether an approval record exists.
How to verify the setup
- Enter the hypothetical sample source text and the draft sentence, then compare whether "authorization" and "validity period" in the draft sentence have corresponding fields in the source. If the source does not state them, judge the draft as unsupported, record the evidence gap, and do not judge it as a denial or a conflict.
- Enter purchase order CG-20250310-07 and compare the items returned by full-text retrieval and by semantic retrieval. Observe whether full-text retrieval hits that number; if it does not, record the matter as pending verification and do not infer that the data is missing.
- Enter a clearly labelled paper empty result [] and check whether the downstream node continues to execute. The knowledge base search node outputs a citation array whose length may be 0, and the output path continues even when nothing is found. Observe whether downstream handles it as an empty array, without claiming that a fictional rule always returns [] on an empty search.