What the documents look like
Complaint and regulatory reply work is addressed to a platform or a regulator, so the materials must carry formal evidence and remediation text. Errors in this category can draw administrative investigation, platform penalties, or brand exposure. Daily documents include complaint tickets, regulatory inquiry letters, order evidence, and remediation reports. These materials commonly arrive from the 12315 platform, the platform complaint center, customer service chat, and the order system, and their formats and wording conventions are not uniform. The comparison task therefore begins by mapping each proposed sentence to the source text it cites, then judging whether that citation can support the wording. Data updates are irregular and case-by-case; this page does not cover batch synchronization or transaction-triggered processing.
A citation-tracing setup has to keep three things separate: what the retrieval step returned, what a human checked against the original, and what a downstream system still has to verify. Retrieval returns candidate content. It does not confirm authenticity, grant permission, or complete any business approval.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model, for vector search. Its output is a citation array, and the array length may be zero, which means the output path still runs even when nothing was retrieved. Citation content is handed to the answering step, where each answer sentence is checked line by line against the cited source text. Similarity and minimum relevance affect recall ordering only; they do not act as permission or entity isolation. Retrieval is not authorization, and it does not confirm that a material is genuine.
Knowledge bases with different vector models cannot be associated together. Check the model identifier first and exclude any base that does not match. Mixed retrieval fuses multiple result paths through RRF, which is a result-fusion process rather than a switch. Image retrieval requires image input and a model that supports it. Reranking requires a rerank model and must be enabled in the settings. The optional paths in the overall retrieval flow are not all executed on every request.
Assume a paper exercise sample. The source material text reads: "Ticket number A-001. The complainant placed an order on 5 March 2024, order amount 200 yuan, applied for a refund on 8 March 2024, and the platform completed the refund on 9 March 2024." The proposed sentence reads: "This order has completed a refund, and the complainant's claim of non-refund is inconsistent with the record." This sample is used to check subject, object, status, and validity scope. It does not refer to a real customer or a lawful identity number.
Configuration and implementation checks
| Configuration item | Use on this page | How to check |
|---|---|---|
| Associated knowledge bases | Limit the material scope for this comparison | Check the selected knowledge bases and material version, and confirm the vector model identifier matches |
| Search mode | Choose semantic, full-text, or mixed retrieval by material type | Use full-text or mixed retrieval for ticket and number material, and check the matched source text |
| Minimum relevance | Filter low-relevance candidates | For this exercise, start at 0.3 and compare with 0.5, observing how candidates change |
| Citation limit | Control the total cited tokens | For this exercise, start at 2000 and compare with 4000, observing how the number of citations changes |
| Citation content output | Carry the citation array | Check whether the array is passed to the answering step and compared line by line with the source |
| Citation display check | Verify the displayed source text | Use a traceable sample to check that displayed content matches the source |
The starting values for minimum relevance and citation limit are exercise values chosen for comparison against another value. They are not official defaults. Thinking budget is a model thinking-intensity setting and is unrelated to the citation token limit.
Three common errors
- Treating a retrieval hit as authorization or as factual confirmation. Retrieval returns candidate content only. Business verification, approval, and database writes require downstream implementation. A pending review does not show that an operation permission is absent.
- Forcing knowledge bases with different vector models into a joint association. The documentation states that only knowledge bases with the same vector model can be selected together. Check the identifier first, exclude any base that does not match, and do not run cross-model retrieval to compare missed recalls.
- Judging records from different fields or different periods as contradictory. Only an explicit mutually exclusive value in the same subject, same period, and same field counts as a contradiction. A record that does not state a fact, or that shows only a related stage, is unsupported or pending verification.
How to verify the setup
- Enter the exercise sample source text and the proposed sentence, then compare whether the proposed sentence "has completed a refund" and the source text "the platform completed the refund on 9 March 2024" share the same subject, same period, and same field. When the two agree, the condition is met. If the source text states only that a refund was applied for, the claim is unsupported.
- Enter an empty citation array []. Compare whether the downstream answering step still runs and records the gap. When the output path runs without error, the condition is met. Do not claim that a rule-based retrieval will necessarily return [].
- Enter two knowledge bases with different vector model identifiers and compare whether they are excluded. When the mismatched identifier is excluded and the actual observation is recorded, the condition is met. Do not infer the runtime result of a joint association.
These three checks describe what to compare and what would satisfy the condition. They do not report that any test was executed, that a deployment was approved, or that a business action occurred. Where a document field is not supplied, the value remains unknown, and the gap is recorded rather than filled in.