What the documents look like
Livestream after-sales work tends to arrive in bursts after a single high-traffic session. Refund requests, replacement orders, and explanation commitments pile up within a short window, and the customer service team drafts product descriptions and marketing copy from four recurring materials: livestream order lists, refund applications, replacement orders, and script templates. Each drafted sentence about a promise, an order status, or a replacement arrangement needs to point back to the specific material and the specific version it came from. When a drafted sentence does not match the underlying material, the consequences are concentrated negative reviews, platform intervention, and damage to creator partnerships. The update rhythm is per-session and daily, which means the material versions a draft relies on can shift between the moment a promise was made and the moment the copy is written.
Constraints at the citation tracing step
The 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, and that array may have a length of zero, which means the downstream chain still runs even when nothing is retrieved. The search scope is first checked against the selected knowledge bases and the material versions they contain. Retrieved results then pass to the answering step, where each drafted sentence is compared line by line against the cited original text.
Semantic search, full-text search, hybrid search, and reranking are optional retrieval paths. Image retrieval requires corresponding image input and model capability, and reranking requires a rerank model enabled by configuration. The minimum relevance score takes effect only under semantic search or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. Knowledge bases built on different vector models cannot be associated together; the identifier is checked and non-matching bases are excluded.
Hypothetical paper exercise sample: the material original text reads "Order A from this livestream session submitted a refund application on March 1, 2026, status pending review," "Replacement order B corresponds to Order A, replacement quantity 1 item, warehouse has not yet shipped," and "Script template C states that the replacement time limit is based on the logistics trace after warehouse shipment." The draft sentence to be judged reads "Order A's refund has been approved, and replacement order B has shipped." When this sentence is checked, the material shows only that the refund is pending review and the replacement has not shipped, so the draft sentence lacks support.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Select the after-sales material base and the script base that share the same vector model | Check whether the vector model identifiers of the selected bases match; exclude bases with different identifiers |
| Search mode | Hybrid search | Compare with semantic search alone and observe whether order-number and ticket-number queries are hit more readily |
| Minimum relevance score | Hypothetical starting point 0.5 for this exercise | Compare with the displayed value and observe whether filtered citations still cover the material original text |
| Citation limit | Hypothetical starting point 3000 tokens for this exercise | Compare with a smaller value and observe whether the material original text is truncated |
| Citation output handoff | Pass the citation array to the answering step | Check whether each drafted sentence can be traced back to the cited original text |
| Material version check | Use one current version and one historical version of the material | Compare the recorded values for refund status and replacement status across the two versions |
Three common errors
- Writing "pending review" as "approved." The input is the material original text "refund application, status pending review," compared with the draft sentence "refund has been approved." The observable result is that the draft sentence exceeds the scope of the material.
- Writing "not yet shipped" as "has shipped." The input is "warehouse has not yet shipped," compared with the draft sentence "has shipped." The observable result is that the status field does not match.
- Associating knowledge bases that use different vector models. The input is the vector model identifiers of two bases, compared for consistency. The observable result is that a base with a different identifier is excluded, and no cross-model retrieval comparison is performed.
How to verify the setup
- For the sample material original text and the draft sentence, the input is the three original sentences and one draft sentence. Compare the refund status field and the replacement status field. The condition is satisfied only when the draft sentence matches the material; the expected result for this sample is not satisfied.
- For the empty citation handoff, the input is the paper citation array []. Compare whether the downstream answering step still runs. The condition is satisfied only when it runs through without error.
- For material version currency, the input is the current version and the historical version of the material. Compare the values of the same field. When the values differ, record the difference and keep the status as pending verification, without judging it as a negation or a conflict.