What the documents look like
Livestream after-sales refund requests tend to arrive in a burst after a single traffic event ends, and they are strongly tied to what the host said during that event. Store service and internal training work with four document types: livestream order lists, refund applications, reshipment orders, and script templates. These materials come from platform after-sales records, customer service IM, warehouse WMS, and logistics tracking, and they update after each session and daily. The task here is citation tracing: during training and handover, staff must be able to state which material a given explanation or document came from, which version it is, and what time range it covers. When material versions differ, the refund conclusion for the same order can differ, so checking the source comes before unifying the script.
Constraints at the citation tracing step
A knowledge base search node can associate one or more knowledge bases that share the same vector model. Its output is an array of cited content whose length can be zero. A zero-length output means the search path completed; it does not mean the workflow terminated with an error. Whether retrieved results are passed to the answering step, and whether the answer corresponds line by line to the cited source text, requires manual checking. Retrieval itself does not constitute authorization, and a similarity score does not perform entity or permission filtering. Hybrid search merges multiple result paths through RRF. Reranking requires a model and must be enabled according to settings. Image retrieval requires corresponding image input and model capability. These paths are not all executed on every request.
Hypothetical drill sample: one paper material from a store after-sales team states: "For the period March 1 to March 7, 2025, refund application for order A1001 was registered as refundable by customer service, based on the host's statement during the March 1 livestream that 'seven-day no-reason return is available'; warehouse reshipment order B2002 corresponds to the same order and is pending outbound." A second material states: "Refund application for order A1001 is under review; refund is not yet complete." The draft sentence to be judged is: "Order A1001 refund is complete and the reshipment has been sent." When checking, compare the subject, order number, time range, and status fields in both materials. The first material supports only "registered as refundable" and "reshipment pending outbound." The second material supports only "under review." The draft sentence's "refund is complete" and "has been sent" are not supported by either material and must remain pending verification.
Configuration and implementation checks
| Configuration or check item | Purpose on this page | Basis and comparison method |
|---|---|---|
| Associated knowledge bases | Limit the retrieval scope for training materials | Associate only knowledge bases with the same vector model; verify identifiers first and exclude those that do not match |
| Cited content output | Receive retrieval results | Output is an array whose length can be zero; check downstream design against a paper input of [] |
| Citation limit | Control the amount of cited content per search | Measured in tokens; set a hypothetical starting point based on actual material length and answer needs, then compare |
| Minimum relevance | Filter low-relevance results | Takes effect only in semantic search or when reranking is used; adjust based on actual returns |
| Search mode | Match material type | Use full-text or hybrid search for numbers and order IDs; use semantic search for natural-language explanations |
| Answer-to-citation check | Trace back to source text | Compare each answer sentence against cited source text on subject, order number, and status fields |
The hypothetical trial starting point for the citation limit is 800 tokens. The reason is that the two sample material sentences total roughly one hundred characters, leaving room for merging several materials. Compare against 400 tokens and 1600 tokens to observe whether citations are truncated or irrelevant fragments are introduced. This value is a hypothetical starting point for this page, not an official default.
Three common errors
- Treating a retrieval hit as business authorization. Input is a refund application material; compare the retrieval output against the authorization status field. The observable result is that cited content exists while the refund status remains under review. Retrieval does not execute a refund.
- Treating an empty citation array as a process failure. Input is a query with no matching material; compare the node output against the downstream branch design. The observable result is an output of [] with the path continuing. Whether an error is reported depends on actual configuration and cannot be assumed.
- Forcing association of knowledge bases with different vector models. Input is two knowledge base identifiers with different vector models; compare the model identifier fields. The observable result is that the identifiers do not match. Exclude the non-matching one and record it; do not run cross-model retrieval to compare missed recalls.
How to verify the setup
- Use the first hypothetical material's original text as retrieval input. Compare whether the returned citation contains the original wording "registered as refundable" and "reshipment pending outbound." The condition is met when the cited sentence matches the material on subject, order number, and time range.
- Compare the draft sentence "Order A1001 refund is complete and the reshipment has been sent" field by field against both materials. The condition is met when the check can point out that "refund is complete" and "has been sent" are not stated in either material and remain pending verification.
- Use a paper input of [] to check the downstream receiving design. Compare whether the output path continues. The condition is met when the path is observed to complete and the empty array is not judged as a business failure. Actual error behavior depends on configuration observation; what has not been observed remains unknown.