What the documents look like
Livestream after-sales work clusters around individual broadcast sessions. Refund applications, reshipment orders and script templates tend to arrive together after a session ends, and the follow-up questions in customer service IM cross with feedback from the warehouse exception desk. The brief lists the main data sources as the shop after-sales console, customer service IM, the warehouse WMS and logistics tracking. The documents handled day to day are livestream order lists, refund applications, reshipment orders and script templates. Updates are described as per session and daily, so the knowledge base that retrieval depends on needs to stay aligned with the same day's document wording.
The stated difference from ordinary platform after-sales is that livestream after-sales breaks out after a single session's traffic and is strongly tied to what the host promised. Promise wording is often scattered across script templates and customer service IM records. When retrieval returns only generic phrasing, the agent lacks a checkable source text for explaining the promise consistently. Identifying surge refunds, arranging missed or incorrect reshipments, and explaining promises consistently are the three most time-consuming steps in this category, and retrieval directly affects the quality of response in the first two.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. During implementation, check the vector model identifier first, exclude the ones that do not match, and record the association result actually observed. The retrieval step itself does not perform authorization checks. A similarity score is not a permission or entity filter, and the scope of the retrieved knowledge base is not the same as the scope of operating authority.
Full-text retrieval suits exact queries such as numbers, models and proper nouns. Semantic retrieval suits natural-language questions. Hybrid retrieval merges the two paths through RRF. RRF is a fusion process, not a separate switch that can be turned on. Reranking requires a ReRank model and must be enabled according to the settings. Image retrieval requires the corresponding image input and model capability. Reference content is output as an array whose length can be zero, and the empty-result path still runs through.
Hypothetical drill sample: after a livestream session on 12 March 2025, customer service receives a reshipment application. The source material reads: "Order A-2031, the customer reports that the outer packaging of the goods received on 10 March was damaged, and applies for one reshipment of the same item; the warehouse WMS shows the order has been shipped out." The draft sentence to be assessed is: "The reshipment for this order has been completed and the tracking number has been synchronized." In this sample the reshipment order is at the application stage, and the warehouse record shows only that it has left the warehouse. The completion of reshipment and the tracking number synchronization claimed in the draft sentence have no corresponding field support in the material.
Configuration and implementation checks
The table below lists, around the sample above, settings already described in the official material or implementation check items relevant to this feature and the primary task. The values are starting points for this hypothetical test, used to compare against other values, and are not official defaults.
| Configuration item | Starting point for this hypothetical test | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge base | Associate one after-sales script library that uses the same vector model | Check the vector model identifier first, then compare with another candidate library whether they can be associated together |
| Search mode | Test semantic retrieval first, then hybrid retrieval | Use the same test question and observe the difference in returned content |
| Reference limit | Set the starting point to a token amount that can hold one complete script template | Test with actual material of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Set the starting point to a middle value | Label relevant and irrelevant samples first, then compare the filtering result |
| Question optimization | Enable for multi-turn follow-up scenarios, disable for single-turn explicit questions | Compare the original question with the optimized question and check whether the subject has been altered incorrectly |
| Result reranking | Enable when there are many candidates, and check the rerank score | Compare ordering before and after reranking, and confirm whether the score is used for filtering |
Three common errors
- Treating a retrieval hit as authorization. The input is one refund application, and retrieval returns a related script. If this is taken to mean the refund has been approved, the comparison is between the application status field and the approval field, and the observed result is that the two values do not match. The retrieval result does not contain approval status and cannot replace business approval.
- Treating a completion statement in a draft as fact. The input is the draft sentence "The reshipment for this order has been completed" from the sample above. The comparison is between the current status of the reshipment order and the logistics tracking field. The observed result is that the material shows only that the order has left the warehouse, with no record of receipt or completion. The draft sentence is unsupported.
- Associating knowledge bases across vector models. The input is two knowledge bases with different vector model identifiers. The comparison is between the two vector model identifiers, and the observed result is that they do not match. The non-matching one should be excluded and recorded, without running cross-model retrieval to compare missed recalls.
How to verify the setup
- Check the retrieval scope against the sample source text. The input is "Order A-2031, applies for one reshipment of the same item." The comparison is whether the reference content returned by retrieval includes the reshipment conditions and status field for that order. The observed result is that the reference content covers the application-stage information and does not contain a completion statement, which satisfies the check condition.
- Check the judgment boundary against the draft sentence. The input is the draft sentence "The reshipment for this order has been completed and the tracking number has been synchronized." The comparison is whether the material contains fields for completed reshipment and tracking number synchronization. The observed result is that the material does not state such fields, so the judgment is pending verification and not a conflict.
- Check downstream handling with an empty result. The input is a clearly paper-based empty array. The comparison is the downstream node's design for handling reference content. The observed result is that the empty-result path runs through and downstream can recognize the absence of reference content. No claim is made that retrieval under a fictional rule will necessarily return an empty array.