What the documents look like
Complaints and official replies are directed at platforms or supervisory authorities, so the material must carry formal evidence and remediation wording. The working sources are the 12315 platform, platform complaint centers, customer service chat, and the order system. Day-to-day handling covers complaint tickets, regulatory inquiry letters, order evidence, and remediation reports. These materials change per case rather than on a fixed cycle, and a single matter moves between the ticket, the inquiry letter, the evidence package, and the reply draft. That shape matters at retrieval time: the search step has to keep tickets, inquiry letters, order evidence, and remediation reports distinguishable, and any reply wording has to be traceable back to a specific piece of evidence. A retrieval result that mixes a remediation report into an evidence question, or that returns wording without its supporting record, creates rework before the reply is sent.
Constraints at the retrieval step
In FastGPT, a knowledge base is made of libraries, collections, and data. The smallest search unit is the library; collections only classify data and do not affect search results. 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; the correct handling is to check the model identifier and exclude the ones that do not match, without running a cross-model search to compare what is missed.
A single retrieval combines question optimization, semantic search, full-text search, or hybrid search according to the input content and search parameters, then merges results through RRF, applies ReRank for a second ordering, filters by similarity, and trims by the reference limit. Semantic search depends on vector distance. Full-text search suits numbers, model names, and proper nouns. Hybrid search merges both through RRF. The minimum relevance value takes effect only in semantic search or when result reranking is used, and it is a filter threshold rather than a permission or entity isolation control. The reference limit is measured in tokens. The node outputs reference content as an array whose length can be zero, which means the output path still runs even when nothing is found.
A hypothetical paper exercise: a shop customer service supervisor receives a ticket transferred from a platform. The ticket body reads: "The consumer reports that order A arrived signed-for with visible damage to the exterior, and requests a refund plus a written handling statement." The supervisor drafts a handover request: "Please confirm the signed-for status of order A and where responsibility for the damage lies, then issue a refund reply on that basis." This sample contains a subject (the consumer and the shop), an object (order A), a state (reported damage after signing, refund demand), and a scope (handling statement and reply). The sample is hypothetical and points to no real customer or real order.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the complaint ticket and order evidence libraries that share one vector model | Check the vector model identifier of each library, exclude mismatches, and do not run a cross-model search |
| Search mode | Start by testing semantic search on natural-language ticket and inquiry questions | Compare candidate content for the same test question against full-text and hybrid search |
| Reference limit | Hypothetical starting value of a few tokens, enough for one ticket body plus one evidence item | Test with actual materials of different lengths and see whether the reference content stays complete, then adjust |
| Minimum relevance | Hypothetical starting value as a 0-1 number | Label relevant and irrelevant samples first, then compare the filtered results at that threshold |
| Question optimization | Enable for multi-turn follow-ups and short questions; compare the original and optimized question | Check whether subject and object were altered; a clear question can be left without it |
| Result reranking | Used when a rerank model is available and enabled by the setting | Compare rerank scores against the unranked order and confirm the candidate pool is sufficient |
Three common errors
- Treating semantic search or the minimum relevance value as a permission filter. Similarity is a number and does not separate business entities or authorization scope; retrieving material does not mean the material may be used in an outbound reply. The check is to confirm the source and access right of the material, and to record the gap when that is missing.
- Treating an output array length of zero as an anomaly. The documentation states that the reference length can be zero and the path still runs. The check is to feed a clearly labelled paper input of an empty array and verify the downstream design against it, without claiming that a fictional rule search must return an empty array.
- Treating a pending review, a quotation validity period, or a draft sentence as a completed fact. A pending review does not show that an operation permission is missing; a quotation validity period does not state a supply period; a draft sentence is not an issued reply. The check must hold subject, period, and field constant.
How to verify the setup
- Enter the hypothetical ticket text above together with the draft sentence, and compare whether the retrieved reference content covers order A, the signed-for status, and the damage demand. The expected observation is candidate content sharing the subject and object of the ticket. If the returned content carries a different subject or order number, record it as unsupported and keep the gap.
- Enter the same test question through semantic search and through full-text search, and compare the candidate content. The expected observation is that number-type queries hit more directly in full-text search. If neither mode returns a candidate, record the empty result and check the downstream design against an empty array, without inferring a failure cause.
- Enter the draft sentence and compare whether the optimized question retains order A and the signed-for status. The expected observation is that subject and object are not rewritten. If the order identifier is lost after optimization, record the difference and keep the original question. All three are paper checks; no retrieval has been executed and no submission has been made.