What the documents look like
Complaints and regulatory replies are directed at platforms or regulators and require formal evidence and rectification language. Source materials include the 12315 platform, platform complaint centres, customer service chats and order systems. Day-to-day documents are complaint tickets, regulatory inquiry letters, order evidence and rectification reports.
For product information and selection questions, the work at this step is to bind product materials, order evidence and reply text to checkable sources so that each cited item can be traced back to an original document. Cases arise irregularly, on a per-case basis, and material versions can differ between cases. The same product may need separate verification of its material version in each case.
A complaint ticket may state that the complainant asserts order A was signed for in March 2025 and that the product page listed the material as pure cotton. A merchant reply may state that the batch page showed a blended fabric and attach a February 2025 product detail screenshot. A regulatory inquiry letter may ask when the material description for that product changed. These three items form the material set for the sample below.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model and outputs a citation array whose length may be zero. Even when no content is retrieved, the output path still runs. Citation content passed to the answering step must be compared item by item against the source text; retrieval results do not constitute authorization and do not establish that material is authentic.
A similarity score reflects content relevance only. It does not act as a permission or entity filter, and it does not exclude unrelated businesses or unauthorized records. Hybrid retrieval merges multiple result paths through RRF, and result reranking uses a ReRank model to produce a 0–1 score. The minimum relevance setting takes effect only under semantic retrieval or when result reranking is used. RRF is a fusion process for hybrid results, not a switch that is turned on.
Hypothetical paper exercise sample. The materials to be checked are three sentences: (1) the complaint ticket body, which states that the complainant asserts order A was signed for in March 2025 and that the product page listed the material as pure cotton; (2) the merchant reply, which states that the batch page showed a blended fabric and attaches a February 2025 product detail screenshot; (3) the regulatory inquiry letter, which asks when the material description for that product changed. The draft sentence to be assessed is: "The product page material description changed from pure cotton to blended fabric starting in February 2025." The draft sentence asserts a change time. The available materials show only that the February 2025 screenshot carried a blended fabric description. They do not record the prior version or the change action itself. The draft sentence is therefore unsupported on the change time and requires the prior version or a change record before it can be assessed.
Configuration and implementation checks
| Configuration or check item | Use on this page | How to check |
|---|---|---|
| Associated knowledge bases | Select the knowledge bases holding product materials and order evidence | Review the selected knowledge bases and material versions, and confirm the vector model identifiers match |
| Citation limit | Controls the tokens of content cited per search | For this hypothetical exercise, set a small starting value chosen because the sample materials are short, then compare against total sample length and adjust |
| Minimum relevance | Filters low-relevance results | Takes effect only under semantic retrieval or result reranking; observe citation counts before and after filtering using the sample question |
| Search mode | Choose semantic, full-text or hybrid retrieval | Observe hits separately for material-name queries and order-number queries |
| Result reranking | Re-orders candidate text a second time | Requires a ReRank model enabled in settings; compare ordering before and after reranking |
| Citation output | Passes the citation array to the answering step | Use the sample question to check whether the output array corresponds to the source statements in the reply |
Three common errors
- Treating a similarity score as a permission or entity filter. Input: product materials belonging to different stores inside one knowledge base. Compare the score values against material ownership and observe whether a high-scoring item whose ownership does not match is still recalled. The score does not perform isolation; ownership must be checked separately.
- Treating retrieval output directly as the reply's source. Input: the sample question. Compare the citation array text against the draft sentence and observe whether the draft sentence introduces a change time that the materials do not record. When the citation array length is zero, the downstream path still runs, so this must be checked separately.
- Associating knowledge bases that use different vector models. Input: two knowledge base identifiers. Compare their vector model identifiers and observe whether a mismatch is excluded. Cross-model association is not permitted, and no cross-model retrieval comparison is performed.
How to verify the setup
- Use the sample complaint ticket body and the merchant reply screenshot description as input. Compare the draft sentence "The product page material description changed from pure cotton to blended fabric starting in February 2025" against the source text, and observe whether the materials record the prior version or the change action. When they do not, record the result as unsupported and retain the evidence gap.
- Use the sample inquiry letter requirement as input. Compare each source statement in the reply against the entries in the citation array, and observe whether each source statement can be mapped to a specific material. Record entries that cannot be mapped as pending verification.
- Use two knowledge base identifiers as input. Compare the vector model identifiers and observe whether a mismatched identifier is excluded from the association scope. Record the actual observation after exclusion without inferring runtime results.
For a zero-length citation array, a paper input of an empty array can be used to check the downstream design. This does not assert that any constructed rule will return an empty array.