What the documents look like
Complaints and official replies are addressed to a platform or a regulatory authority, so the material must carry formal evidence and remediation wording. Sources include the 12315 platform, platform complaint centres, customer service chat and the order system. The working documents are complaint tickets, regulatory inquiry letters, order evidence and remediation reports. Records are updated irregularly, case by case. For retrieval, the practical consequence is that records for one ticket sit in different places at different stages. A handover must separate what has already been done, what is a configurable field, and what remains a manual check.
A complaint ticket, an inquiry letter, order evidence and a remediation report are distinct document identities. They describe different objects and different stages. A ticket number identifies a case; it does not identify an approval. A remediation report is a document that may or may not have been submitted. Retrieval can return any of these as reference content, and reference content is not a statement about the status of the case.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Check the vector model identifier of each candidate base first and exclude those that do not match; do not run cross-model retrieval to compare results.
One retrieval combines problem optimisation, semantic retrieval, full-text retrieval, hybrid retrieval, image retrieval and reranking according to the input and the search parameters. These paths run conditionally. Image retrieval requires an image input and the relevant model capability. Reranking requires a rerank model and must be enabled in the settings. The full retrieval flow does not execute every path on every request.
Reference content is output as an array and its length can be zero. A zero-length result still lets the downstream chain proceed. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold, not a final ordering rule, and it is not a permission or entity filter. A retrieval hit is not authorisation, approval, signature or contract effect.
Hypothetical paper exercise sample. A shop customer service supervisor hands over a complaint ticket to a legal compliance colleague. The source text reads: "Ticket number A-017, complaint target is this shop, complaint matter is delayed shipment, current status is pending review. The regulatory inquiry letter requires a remediation reply to be submitted within ten working days of receipt." The draft sentence to be assessed reads: "This ticket has been approved by the platform and the remediation reply has been submitted." The check point: the source text states only that the status is pending review. It does not state approval and does not state that a remediation reply was submitted. Pending review does not show that an operating permission is absent, and both approval and submission remain unsupported.
Configuration and implementation checks
| Configuration or check item | What it means for this page | Hypothetical starting point and comparison method |
|---|---|---|
| Associated knowledge bases | Select complaint ticket, regulatory inquiry letter, order evidence and remediation report bases that use the same vector model | Check each base's vector model identifier first; exclude any that do not match; do not run cross-model retrieval |
| Search mode | Compare semantic, full-text and hybrid retrieval against the same question set | Test identifier-style questions and natural-language questions against the same material |
| Reference upper limit | The unit is tokens | Test with material of different lengths to see whether content stays complete; state the reason for the starting value and compare it with other values |
| Minimum relevance | Takes effect only under semantic retrieval or when result reranking is used | Mark relevant and irrelevant samples first, then adjust the threshold |
| Problem optimisation | Coreference removal and question expansion | Compare the original question with the optimised question and check whether the subject was changed incorrectly |
| Empty-result handling | Reference content length can be zero | Use the paper input [] to check the downstream design; do not infer that a rules-based retrieval must return [] |
The reference upper limit is expressed in tokens rather than a top-k count because chunk lengths differ across mixed knowledge bases. Any starting value used in this exercise is a hypothetical trial value with a stated reason, not an official default. The minimum relevance value is a number between 0 and 1 that filters low-relevance results; it does not exclude unrelated businesses or unauthorised records.
Three common errors
- Treating a ticket pending review as approved. Input: the sample source text "current status is pending review". Compare whether an approval field is stated. Observable result: approval is unsupported. Pending review is not evidence that an operating permission is missing.
- Treating a retrieval hit as a permission or an authorisation. Input: one item of reference content. Compare whether an authorisation record exists. Observable result: retrieval does not create authorisation, approval, signature or contract effect.
- Forcing knowledge bases with different vector models into one association. Input: the vector model identifiers of two bases. Compare whether the identifiers match. Observable result: bases whose identifiers do not match must be excluded. Do not then assert a runtime consequence such as missed recall or redundant returns; record only the identifier check and the actual observation.
How to verify the setup
- Input the sample source text and the draft sentence. Compare whether the approval and submission fields are stated. The condition is met when the draft sentence is marked unsupported. No claim is made that this test has been run.
- Input identifier-style questions and natural-language questions. Compare the reference content returned under each search mode. The condition is met when the differences between modes can be described and no guarantee is made that the first item is correct or that all relevant items are recalled.
- Input the paper empty result []. Compare the downstream handling design. The condition is met when it can be shown that a zero-length array can pass through and no claim is made that a rules-based retrieval must return [].
Across all three checks, keep the subject, period and field aligned. Only mutually exclusive values in the same field count as a contradiction. A field that is not stated, or evidence that shows only a related stage, counts as unsupported or pending verification. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where the material needed for a check is not present in an accessible retrieval scope or a defined manual input, record the gap and leave it open.