What the documents look like
Seller ratings and experience-score maintenance in platform e-commerce turns on consumer feedback that arrives after fulfilment and after-sales service. The working materials are poor-review handling records, experience-score detail sheets, service diagnostic reports and rectification ledgers. These materials update daily or weekly. Copy drafting for product and marketing text needs quotable wording drawn from those materials: the attributed cause recorded for a specific negative review, the status recorded for a rectification item, the judgement a diagnostic report makes about a service stage. The subject, object, status and validity scope of a statement are often spread across separate documents, so the retrieval step returns candidate content and a human then checks whether a proposed sentence matches the source wording.
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 vector model identifier is checked first and non-matching bases are excluded. Search modes include semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a switch that can be turned on. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it is a filtering threshold. The reference cap is measured in tokens. Question optimisation performs coreference resolution and question expansion and adds one model call. Reference content is output as an array and its length can be zero.
Hypothetical paper exercise sample. Source text one: "From 1 March 2025 to 31 March 2025, the after-sales review handling record for a given store shows that the negative-review attribution conclusion for order A is logistics timeliness, and the rectification status is supplementary evidence submitted, pending platform review." Source text two: "The service diagnostic report for the same period states that the customer service response stage of this store is not listed in the current rectification ledger." Proposed draft sentence to be judged: "This store has completed all rectification of the logistics timeliness issue."
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only bases on the same vector model that contain the two source texts above | Check the vector model identifier first and exclude bases on other models |
| Search mode | Hybrid retrieval | Compare recalled content against semantic retrieval alone and full-text retrieval alone |
| Reference cap | Assumed starting value of 2000 tokens | Test with material fragments of different lengths for content completeness, then adjust |
| Minimum relevance | Assumed starting value of 0.3 | Label relevant and irrelevant samples first, then compare filtering behaviour |
| Question optimisation | Enabled | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enabled according to configuration | Compare candidate order before and after reranking and confirm the model is available |
Three common errors
- Treating the minimum relevance value as a permission or entity filter. That value only filters low-relevance results and does not exclude unrelated businesses or unauthorised records. Check: enter the proposed draft sentence, compare whether the recalled content still contains source text one and source text two, and observe whether records unrelated to the subject appear. A relevance score is not a category, entity or permission filter, and retrieval is not authorisation.
- Reading a pending-review status as completed rectification. Source text one records the rectification status as supplementary evidence submitted, pending platform review, while the proposed sentence states that all rectification is complete. Check: compare the proposed sentence with source text one field by field and observe whether "pending platform review" and "completed" occupy mutually exclusive values on the same field. A pending request is not an approval.
- Reading "not listed in the ledger" as a negative conclusion. Source text two states only that the customer service response stage is not listed in the current rectification ledger; it does not record whether that stage has a problem. Check: compare whether the proposed sentence contains a negative statement about that stage and observe whether the material provides a corresponding field. Missing evidence remains missing.
How to verify the setup
- Enter the proposed draft sentence "This store has completed all rectification of the logistics timeliness issue." Compare the recalled content against the rectification status field in source text one and observe whether the recalled result contains the wording "pending platform review." If the recalled content does not contain that field, record the sentence as unsupported rather than as contradicted. No first result is guaranteed and no full recall is guaranteed.
- Enter the proposed draft sentence and compare whether the recalled content contains the statement in source text two about the customer service response stage. Observe whether the proposed sentence makes a judgement about that stage that the material does not record. If the material provides no such field, record the point as awaiting verification. A relevance threshold does not exclude unrelated businesses or unauthorised records, so this comparison is about wording support, not about access control.
- Use a paper input of an empty array to check the downstream design, and observe whether the downstream path runs when the reference content length is zero. This check does not assert that a fictional rule search returns an empty array; it verifies only how an empty result is carried forward. The documented behaviour is that reference content is output as an array whose length can be zero, so the output path still runs even when nothing is retrieved.
Across these checks, distinguish what the node already does, which fields are actually configurable, which steps are manual comparisons, and which business actions would require separate downstream implementation. Forms collect values; business verification, approval, database writes and external synchronisation need their stated downstream implementation. Scheduled triggers and hidden controls are not assumed. Image retrieval requires an image input and model capability, and reranking requires a model enabled according to configuration; the optional paths in the overall retrieval flow are not executed on every request. Where a comparison depends on a document field that was not provided, that field remains unknown.