What the documents look like
Livestream after-sales work concentrates after a single broadcast: refund requests and reshipment orders arrive in a burst, and the after-sales wording has to stay consistent across the whole batch of orders from that session. The working documents are livestream order lists, refund applications, reshipment orders, and script templates. They come from shop after-sales systems, customer-service IM, warehouse WMS, and logistics tracking, and they are updated after each session and daily.
For product-information and option-selection questions, the retrieval step is what lets a service agent pull a cited passage when answering questions about reshipment or refund wording. Because these materials are tied to what the host promised during the broadcast, retrieval needs to match the order number, the subject of the question, and the scope of the promise precisely. The same batch of orders must be explained under one consistent standard, so a retrieved passage that belongs to a different order or a different promise is a real risk.
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 search mode is compared across semantic retrieval, full-text retrieval, and hybrid retrieval for the test question; no single mode is presumed to fit all materials.
The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds an extra model call. The output reference content is returned as an array, and its length can be zero; even when nothing is found, that output path still completes. Retrieval results do not carry out authorization decisions; the scope of associated knowledge bases and account permissions are separate matters.
Hypothetical drill sample: after a livestream session on 10 May 2026, an after-sales supervisor assembles a set of materials to be checked. The source text has three sentences: "In this session's refund applications, the buyer for order A-0510-017 claims the host promised seven-day no-reason return, and the application reason reads 'size does not match'." "Reshipment order B-0510-004 corresponds to order A-0510-017; warehouse WMS shows it left the warehouse on 11 May, and logistics tracking has not updated." "Script template item 3 states: size issues are handled by reshipment, and refunds require the buyer to send the item back first." The draft sentence to be judged is: "Order A-0510-017 has completed its refund, and the reshipment has been cancelled." Authorization boundary for this request: the supervisor only has permission to check materials and draft wording; the execution status of refunds and reshipments must be confirmed by the after-sales system and the warehouse system. This page does not assert that any execution has occurred.
Configuration and implementation checks
The table below lists, around the sample above, documented settings or implementation check items relevant to this feature and the primary task. The values are starting points for this hypothetical trial, with the reason for choosing them and the comparison method stated together; they are not official defaults.
| Setting item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the after-sales script template base and the reshipment order base | The two bases may be associated together only if they use the same vector model; check the vector model identifier first and exclude any that do not match |
| Search mode | Test semantic retrieval first, then test the same question with full-text retrieval | Compare whether the reference content covers the order number and the promise field under each mode |
| Citation limit | Hypothetical starting point of 800 tokens | Take one passage each from a long and a short refund application and script template, observe whether the reference content is truncated, then compare against a longer limit |
| Minimum relevance | Hypothetical starting point of 0.3 | Mark relevant and irrelevant samples by hand first, then compare whether irrelevant fragments are filtered out at this value |
| Question optimization | Enable for follow-up questions such as "what is the second point"; disable for questions containing a complete order number | Compare the original question with the optimized question, and check whether the order number and subject were altered |
| Result reranking | Enable in script-template retrieval where candidates are numerous | The rerank score can be used for filtering; compare the reference order against the case without reranking |
Three common errors
- Treating a retrieval hit as grounds for execution. The input is the draft sentence "Order A-0510-017 has completed its refund, and the reshipment has been cancelled." The comparison objects are the refund status in the after-sales system and the reshipment order status in warehouse WMS. Retrieving script template item 3 only states that size issues are handled by reshipment; the actual refund and reshipment status is not stated in the materials, so the draft sentence is unsupported.
- Forcing knowledge bases on different vector models into a joint association. The input is two bases with different vector model identifiers. Compare the vector model identifiers of the two bases; when the identifiers differ, joint association is not permitted. Exclude the non-matching base and record the actual observation, without inferring a runtime result.
- Treating minimum relevance as a permission or entity filter. The input is one unrelated business record. Compare whether that record is filtered out for low similarity. Minimum relevance is a filtering threshold and does not perform business isolation; whether an unrelated record appears depends on the retrieval scope and the data itself.
How to verify the setup
- Using the sample source sentence "Reshipment order B-0510-004 corresponds to order A-0510-017; warehouse WMS shows it left the warehouse on 11 May, and logistics tracking has not updated," enter the retrieval question "reshipment order status for A-0510-017." Compare whether the reference content contains the reshipment order number and the warehouse-exit date. If the reference content shows that field without mixing in another order number, the retrieval for this item meets the expectation.
- Using the draft sentence "Order A-0510-017 has completed its refund, and the reshipment has been cancelled" as the object to check, compare the refund status field in the after-sales system with the reshipment order status field in warehouse WMS. The condition is met only when both show execution. When the materials do not state this, record it as pending verification and do not judge it as a negation or a conflict.
- For the empty-reference path, enter a retrieval request constructed on paper and expect an empty array []. Compare whether the output reference content length is zero. When the length is zero, the downstream design needs to accept an empty result. Do not claim that a fabricated rule retrieval necessarily returns [].