What the documents look like
Livestream after-sales work clusters after a single broadcast, when refund applications and replacement orders arrive in the same window. The working documents are livestream order lists, refund applications, replacement orders and script templates. Customer service agents answering fulfillment progress or after-sales commitments need each sentence to point back to a specific source entry.
Citation tracing addresses one question: can a stated order status, refund conclusion or replacement arrangement be matched to a checkable line in the source material. The category differs from ordinary platform after-sales work because the volume is concentrated after one broadcast and the commitments are tied to what the host said on air. Data sources named in the brief are the shop after-sales system, customer service IM, warehouse WMS and logistics tracking. Errors here tend to produce clustered negative reviews, platform intervention and damaged creator cooperation.
The generic brief marks its confidence as approximate, so the timing and document mix stay conditional. Update frequency is described per broadcast and daily; this page does not extend that into batch synchronization, full synchronization, real-time or transaction-triggered mechanisms.
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 does not mean the workflow has errored or stopped; the downstream path still runs. The retrieval output needs to be handed to the answering stage, where each answer sentence is compared against the cited original text.
The minimum relevance value is a filter threshold and takes effect only in semantic retrieval or when result reranking is used. It does not perform permission isolation or entity isolation, and it does not separate unrelated businesses or unauthorized records. Retrieval is not authorization.
Knowledge bases can be associated together only when they use the same vector model. Cross-model knowledge bases cannot be associated together. Checking this means comparing the model identifier of each library and excluding those that do not match; it does not mean running a cross-model search to compare missed recalls.
The overall retrieval flow combines problem optimization, semantic retrieval, full-text retrieval, image description retrieval, image vector retrieval, RRF fusion and reranking, then applies similarity filtering and citation length trimming. The optional paths in that flow are not executed on every request. Image retrieval requires image input and matching model capability. Reranking requires a rerank model and must be enabled in the settings. RRF is a fusion step for mixed results, not a toggle.
Assume a paper exercise sample. The source material reads: "Order A-77021, refund application submitted on 12 May 2025, status pending review; replacement order B-3108 created on 13 May, warehouse has not shipped; the host stated on the 10 May broadcast that seven-day no-reason returns apply, and did not mention which party bears shipping costs." The sentence to be judged reads: "The refund for order A-77021 has been approved, and replacement order B-3108 has shipped."
The refund status field in the draft is mutually exclusive with the source value. The outbound status field in the draft is mutually exclusive with the source value. The shipping cost responsibility field is not stated in the source, so that part remains unsupported. Pending review does not prove that an operation permission is missing. An unsigned termination is not a completed termination, and an offered option is not execution.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Livestream after-sales script library and order rules library | Both libraries use the same vector model; verify the model identifier library by library before associating them together |
| Search mode | Hybrid retrieval | Order number queries need full-text recall, while commitment explanations need semantic recall |
| Citation limit | 3000 tokens as the hypothetical starting value for this exercise | Compare against the combined length of one refund application plus one replacement order source text, and confirm margin remains after coverage |
| Minimum relevance | 0.35 as the hypothetical starting value for this exercise | Compare against the score distribution after reranking is enabled, and observe whether low-scoring fragments are filtered |
| Result reranking | Enabled | When candidates are numerous, use the rerank score for secondary ordering and compare with the unranked order |
| Citation output handoff | Citation array passed to the answering node | Check whether the answer corresponds line by line to the cited source text and whether the downstream path continues when the array is empty |
The 3000-token and 0.35 values are hypothetical starting points for this paper exercise, not documented defaults. The comparison method is stated so the values can be adjusted against observed material length and observed score distribution.
Three common errors
- Writing a pending-review refund application as an approved refund. The input is the source material value "status pending review." The comparison is the refund status field. The observable result is that the draft sentence and the source value are mutually exclusive. The correct handling keeps a pending-verification wording.
- Writing a created replacement order as shipped. The input is the source material value "warehouse has not shipped." The comparison is the outbound status field. The observable result is that the draft sentence and the source value are mutually exclusive. The correct handling marks the warehouse stage as incomplete.
- Extending the host's seven-day no-reason statement into the merchant bearing shipping costs. The input is the source material statement that shipping cost responsibility was not mentioned. The comparison is the shipping cost responsibility field. The observable result is that the source does not record that field, so it is recorded as unsupported. The correct handling preserves the evidence gap rather than supplying an authorization or effective clause.
How to verify the setup
- Input the source material text and the draft sentence "The refund for order A-77021 has been approved, and replacement order B-3108 has shipped." Compare the refund status field and the outbound status field. Observe whether the answer contradicts the source values. If it contradicts them, the citation tracing fails for this sample.
- Input a clearly labelled paper empty result
[]to the downstream answering node. Compare whether the downstream path continues executing. Observe whether a termination error appears. If the path continues and no order conclusion is fabricated, the empty-citation handoff condition is met for this sample. - Check the vector model identifiers of the two knowledge bases. Compare whether the identifiers match. Observe whether a cross-model association exists. If the identifiers do not match, exclude that library and record the actual observation. Do not run a cross-model search to compare missed recalls.
These three checks describe expected conditions for the paper exercise sample. They do not report a completed test, an approval, a refund or any other business action. Where the source material does not state a field, the field stays unknown, and the correct output records the gap rather than resolving it.