What the documents look like
Brand livestream operations place scriptwriting, inventory and after-sales responsibility on the brand team. The working documents are the livestream script, the product explanation card, the shift schedule and the livestream review. Copy drafting for products and marketing checks statements against what these documents already state about the product, the activity conditions and the responsible party. Data sources named in the brief include the Douyin Shop compass, Juliang Qianchuan, livestream comments and transaction details, with updates per session or hourly. Each promotional sentence in a draft needs to be traced back to the script or explanation card to confirm whether the subject, object, state and validity range are covered by the material. Content the material does not state remains unsupported and cannot be filled in from fluent wording.
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 citations whose length may be zero; even when no content is found, that output path still completes. The citation cap is controlled in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Both are filtering and trimming controls; neither performs permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a model and is enabled according to settings. Image retrieval requires the corresponding image input and model capability. Knowledge bases built on different vector models cannot be associated together; check the identifier first and exclude those that do not match.
Hypothetical paper exercise: a brand livestream team's product explanation card reads, "The featured item for this session is Model A thermos, capacity 500 ml, 304 stainless steel inner liner, livestream price 99 yuan, activity period from March 1 to March 3, 2026." A draft sentence reads, "Model A thermos uses a 316 stainless steel inner liner, livestream price 99 yuan, valid long term." Checking field by field: the material states 304 for the inner liner material, the draft states 316, which is a mutually exclusive value in the same field; the activity period is limited to three days in the material, while the draft says valid long term, exceeding the material's validity range. Capacity and price match the material and can be retained.
Configuration and implementation checks
| Configuration item | Setting on this page | Basis and check method |
|---|---|---|
| Associated knowledge bases | Select the base holding the product explanation card and livestream script; confirm the vector model identifier matches | Only the same vector model can be associated together; verify the identifier first and exclude non-matching bases |
| Search mode | Hypothetical starting point for this exercise is hybrid retrieval | Number and model information relies on full-text hits; natural-language questions rely on vector recall; compare against a single-mode result |
| Minimum relevance | Hypothetical starting point for this exercise is 0.3 | This value takes effect only under semantic retrieval or result reranking; compare returned content against higher and lower values |
| Citation cap | Set in tokens, covering the full explanation card | The citation cap controls citation content length; compare against the actual length of the explanation card |
| Citation content output | Receives the citation array from the search node; compare each item against the source text | Array length may be zero; an empty array does not equal a workflow error termination |
| Material version check | Retrieve with the current and historical explanation cards separately | Whether a citation is adopted still requires review against the valid version |
Three common errors
- Treating the minimum relevance or a similarity score as a permission or entity filter. The score reflects only how relevant content is to the question; it cannot exclude unrelated businesses or unauthorized records. The citation scope still requires manual checking against the material version.
- Forcing knowledge bases built on different vector models into a shared association. The document requires the same vector model for shared association. When identifiers do not match, exclude the base and record it; do not run cross-model retrieval to compare missed recalls.
- Treating the thinking budget as the citation cap. The thinking budget belongs to the model's reasoning strength configuration and is a different field from the citation token cap. Mixing them shifts the meaning of the parameter.
How to verify the setup
- Enter the explanation card text and the draft sentence from the hypothetical exercise, and compare the inner liner material field: the material states 304, the draft states 316. Observe whether the setup identifies this as a mutually exclusive value in the same field and flags it for revision.
- Enter the same material and compare the activity period field: the material limits the period to March 1 to March 3, 2026, while the draft says valid long term. Observe whether the setup judges this as exceeding the material's validity range.
- Use the paper input [] to check downstream handling: when the search node outputs an empty array, observe whether the downstream design continues execution, and record the actual observed behavior of that path. Do not claim that a fabricated rule retrieval necessarily returns an empty array.