What the documents look like
Livestream after-sales work clusters after a single broadcast's traffic peak. Refund applications and reshipment orders arrive in the same window, and the main material sources are customer service IM, warehouse WMS, and logistics tracking. Operating reports and exception reviews gather the livestream order list, refund applications, reshipment orders, and script templates by session, then check which material and which version supports each conclusion. Data updates after each session and daily, so the reporting period must align with the time the materials were collected. Mixing refunds from different sessions into one review distorts the picture. The correspondence between host commitments and after-sales explanations depends on comparing script template versions against the order list line by line. When the citation source is missing, a conclusion can only be marked as pending verification.
Constraints at the citation tracing step
The knowledge base search node outputs cited content as an array, and the length may be zero. An empty array means the search path ran but retrieved nothing; it does not mean the workflow terminated with an error. Associated knowledge bases must share the same vector model. Knowledge bases with different vector models cannot be associated together, so the check starts by confirming the vector model identifier of each selected knowledge base and excluding those that do not match. The minimum relevance value is a filter threshold and takes effect only in semantic search or when result reranking is used. It does not perform permission or entity isolation, and retrieved results cannot substitute for an authorization decision. The citation limit is controlled in tokens and serves a different purpose from a thinking budget value; the two cannot be moved between fields. Hybrid search merges multiple result paths through RRF. RRF is a fusion process inside hybrid search, not a separate switch that can be turned on. Image retrieval requires image input and model capability, and reranking requires a model and must be enabled in settings. Optional paths shown in the documented flow diagram do not mean every request executes all of them.
Hypothetical drill sample: the source material reads, "March 12, 2026, Session A livestream order list item 047, buyer applied for a refund, reason is that a promised gift was not shipped; customer service IM record shows the promise came from the March 12 script template V3; warehouse WMS shows the gift was not dispatched." The draft sentence to be judged is, "This refund should be directly determined as merchant responsibility because the host commitment was not fulfilled, and a reshipment should be synchronized." In this sample the subject is Session A order 047, the object is the refund application and the gift promise, the state is that the warehouse did not dispatch, the effective scope is the March 12 script template V3, and the authorization boundary is material checking only, with no refund approval or reshipment execution.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select after-sales libraries that share one vector model | Check the vector model identifier against each selected library and exclude those that differ |
| Search mode | Hybrid search | Compare retrieved content against semantic search and full-text search separately, observing differences on identifier-type and natural-language queries |
| Minimum relevance | Hypothetical starting value 0.5 | Compare filtered citation counts at 0.3 and 0.7, observing whether pending-verification material is wrongly filtered out |
| Citation limit | Hypothetical starting value 2000 tokens | Compare citation completeness at 1000 and 4000, observing whether long refund explanations are truncated |
| Result reranking | Enabled | Compare candidate ordering with reranking disabled, observing whether the script template version ranks near the front |
| Citation handoff | Check whether search output is passed to the answering step | Compare each answer against the cited source text and confirm the downstream path still runs when the array is empty |
Three common errors
- Treating retrieved results as an authorization basis. Input is the refund application in the hypothetical sample; compare the retrieved content against refund approval authority and observe whether wording appears that determines merchant responsibility without approval. Retrieval does not grant authorization, and refund approval must be implemented in a downstream system.
- Associating knowledge bases across vector models. Input is two after-sales libraries with different vector models; compare the model identifiers and observe whether joint association makes the retrieval source indistinguishable. Check the identifiers and exclude those that do not match; do not run a cross-model retrieval to compare missed recalls.
- Treating a thinking budget as a citation limit. Input is a budget value from the thinking configuration; compare its token meaning against the citation limit and observe whether a value such as 16384 is entered into the citation limit field. The two serve different purposes and their field meanings must be checked separately.
How to verify the setup
- Input the source material and the draft sentence from the hypothetical sample; compare the draft sentence "should be directly determined as merchant responsibility and a reshipment should be synchronized" against the source material's state of warehouse non-dispatch and script template V3, and observe whether the draft exceeds what the material supports. The condition is met when the draft states only pending verification and contains no wording of a completed determination or a completed reshipment.
- Input an empty citation array []; compare the downstream answering step's handoff design and observe whether the empty array is treated as an error that terminates the path. The condition is met when the path runs and the downstream marks no retrieval, with no assertion that a rule-based search necessarily returns [].
- Input the March 12 script template V3 from the hypothetical sample and the historical March 10 version; compare the text differences between the two versions and observe whether the citation source points to the effective version. The condition is met when the citation carries a version number, unsupported versions are marked pending verification, and pending review is not treated as approved.