What the documents look like
Complaint and regulatory reply work faces a platform or a regulator, so the materials carry formal evidence and rectification wording rather than routine after-sales repair or reshipment notes. The main sources are the 12315 platform, platform complaint centres, customer service chats and the order system. Day-to-day documents include complaint tickets, regulatory inquiry letters, order evidence and rectification reports. The most time-consuming steps are reconstructing the transaction, assembling the evidence package and submitting the rectification reply. Updates are irregular and handled case by case.
This affects retrieval in a specific way. A single ticket can hold several materials that read alike but differ in subject, order number or status. Retrieval has to keep those materials distinguishable, because a returned passage is a citation, not a verified fact. The search step supplies content; confirming what actually happened, who is authorized to act and whether a reply may be sent are separate downstream matters.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Semantic search relies on vector distance; full-text search suits exact queries such as numbers, models and proper nouns; hybrid search merges both paths through RRF. The minimum relevance value applies only under semantic search or when result reranking is used. The citation limit is measured in tokens and caps how much content one search may cite. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Image retrieval requires image input and model capability; reranking requires a configured model enabled under the settings. Retrieval and similarity scores do not perform permission or entity isolation.
Hypothetical paper exercise sample: a complaint ticket submitted by a store reads, "The complainant states that order A-118, placed on 6 May 2024, showed 'shipped within 48 hours' on the product page, and by 10 May it had still not been dispatched. The store service agent replied on 9 May that the warehouse had been urged to act." A regulatory inquiry letter reads, "Please explain the dispatch time of order A-118 and the basis for the page commitment." A draft sentence awaiting judgement reads, "The order was dispatched on 10 May." Checking requires confirming whether a dispatch completion record exists in the materials; the available text shows only that dispatch had still not occurred on 10 May.
Configuration and implementation checks
| Configuration item | Hypothetical trial value for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Complaint ticket base and regulatory letter base, both using the same vector model | Check whether the vector model identifiers of the two bases match; exclude any that do not |
| Search mode | Hybrid search first | The ticket contains an order number and natural-language description; compare semantic and full-text recall results |
| Citation limit | Hypothetical starting point of 2000 tokens | Test with complaint materials of differing length to see whether cited content stays complete, and compare truncation against a smaller value |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then compare filtering results at different thresholds |
| Question optimization | Enabled for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject and order number were altered |
| Result reranking | Enabled when candidates are numerous | Compare ordering before and after reranking, and confirm the reranking model is configured and enabled under the settings |
Three common errors
- Treating a retrieved ticket passage as an already verified fact. Retrieval returns cited content; business verification is completed separately. In the sample, "still not dispatched on 10 May" and the draft sentence "dispatched on 10 May" share the same subject, the same field and the same period, and their values are mutually exclusive, so the draft sentence is unsupported.
- Treating the minimum relevance value as a permission or entity filter. That value filters low-relevance results only; it does not exclude unrelated businesses or unauthorized records. Checking should confirm that access control is implemented in a downstream system, since the retrieval step does not carry that responsibility.
- Linking knowledge bases that use different vector models and then inferring an operating result. The documentation states only that knowledge bases using the same vector model can be linked. Checking should verify the model identifiers, exclude those that do not match, and record the actual observation without inferring missed recall or redundant returns.
How to verify the setup
- Enter the ticket text and the regulatory inquiry letter text from the hypothetical sample, then compare the draft sentence "The order was dispatched on 10 May" with "still not dispatched on 10 May" in the source text. If the retrieval results contain no dispatch completion record, the draft sentence remains unsupported and the evidence gap must be recorded.
- Enter the order number "A-118" and compare full-text search against hybrid search. If full-text search returns the number while semantic search does not, record the difference between the two paths, adjust the search mode or supplement the index, and do not claim that the first result is guaranteed to match.
- Use a paper input of an empty array to check the downstream design, confirming that the output path still runs when the cited content length can be zero. Do not claim that a fictional rule search necessarily returns an empty array; the ability to handle an empty result must be verified by the downstream implementation.
This page covers only complaints and official replies / store service and internal training / knowledge base retrieval and recall. When handing over operations, the authorization state of materials, actions and requests must be kept clear; retrieval or a form must not be treated as authorization to act.