What the documents look like
Customer-service knowledge-base maintenance covers FAQ entries, script templates, product parameter tables and escalation tickets. It is the preparation of reference material that precedes order-fulfillment and after-sales question answering, and it does not directly approve refunds. The material is updated on a weekly or monthly cycle, and it is maintained jointly by customer-service supervisors, knowledge-base operators and product operations staff. Retrieval and recall determine whether a correct answer can be obtained when a customer-service question is asked. An answer that misses the question can escalate complaints and undermine training.
This page addresses only knowledge-base retrieval and recall within order-fulfillment and after-sales question answering. A retrieval result is not itself an authorization, and it does not mean that a refund, a replacement shipment or a termination has been carried out. The documents involved here are FAQ entries, script templates, product parameter tables and escalation tickets, and the relevant stages are the handling of order-fulfillment and after-sales questions.
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. The operator first checks the vector-model identifier, excludes those that do not match, and records the actual observation. No cross-model retrieval is performed to compare missed recalls.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Semantic retrieval relies on vector distance and suits natural-language questions. Full-text retrieval suits numbers, model designations and proper nouns. Hybrid retrieval uses both paths and merges them through RRF. RRF is a fusion process, not a switch that can be turned on. Reranking requires a model and must be enabled according to the settings. Image retrieval requires image input and the corresponding model capability. The optional paths shown in the overall flow diagram do not mean that every request executes all of them.
The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used, and it is a filtering threshold. Question optimization performs coreference resolution and question expansion and makes an additional model call. The cited content is output as an array whose length can be zero, and the empty-result path still proceeds downstream.
Hypothetical paper exercise sample: the source material reads, "Order A submitted a return request on 5 March 2026; the request status is pending review; the after-sales ticket number is SF-2026-0305-01; the product is a desk lamp of model X1; quantity 1." The draft sentence reads, "Order A's return request has been reviewed and approved, and a refund will be issued on 8 March." Point to decide: whether the draft sentence and the source material agree on the request-status field.
Configuration and implementation checks
| Configuration item | Value or check on this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Associate only order-fulfillment and after-sales knowledge bases that use the same vector model | The node can associate one or more knowledge bases that use the same vector model |
| Search mode | Test semantic, full-text and hybrid retrieval with the hypothetical sample | Compare against the test question; no single mode is presumed suitable for all material |
| Citation limit | Starting point for this hypothetical trial is 800 tokens, compared against 400 and 1200 | The unit is tokens; test with material of different lengths to see whether content is complete |
| Minimum relevance | Label relevant and irrelevant samples first, then compare threshold values | Takes effect only in semantic retrieval or when result reranking is used |
| Question optimization | Check whether the optimized question retains Order A and the after-sales ticket number | Used for coreference resolution and question expansion; check whether the subject is altered |
| Result reranking | Confirm the rerank model is configured and enabled according to the settings | Reranking requires a model and must be enabled according to the settings |
Three common errors
- Treating pending review as reviewed and approved. In the hypothetical sample the source material states "the request status is pending review," while the draft sentence states "has been reviewed and approved." The two are mutually exclusive on the same status field, which is a contradiction. Pending review does not demonstrate a lack of operating authority.
- Treating a retrieval hit as authorized execution. Returning cited content only shows that material was obtained. Business actions such as refunds and replacement shipments require downstream implementation, and retrieval does not constitute authorization.
- Treating minimum relevance as a permission or entity filter. That value is a filtering threshold; it does not exclude unrelated businesses or unauthorized records, and retrieval is not equivalent to authorization.
How to verify the setup
- Input the hypothetical source material and the draft sentence, and compare the request-status field: the source material says pending review, the draft sentence says reviewed and approved, and the two are mutually exclusive, which is judged a contradiction. If the draft sentence only says "has submitted a request," which agrees with the source material, no contradiction is supported.
- Input the hypothetical source material, and compare whether the draft sentence retains Order A, after-sales ticket number SF-2026-0305-01 and model X1. Only when the subject, object and number are all present is the sentence checkable. If any item is missing, record it as an evidence gap and do not judge it as a negation.
- Input the hypothetical source material, and check the citation-limit and minimum-relevance settings: observe whether the returned citations include this order's material. If the returned array is empty, proceed downstream on the basis that the citation length can be zero, and do not claim that a fabricated rule causes retrieval to return empty.