What the documents look like
Brand livestream operations carry their own script, stock and after-sales responsibility, so procurement and supplier records decide whether a product card can state on-hand quantity, supply batch and the party handling after-sales. The daily document surface includes livestream scripts, product briefing cards, shift schedules and livestream reviews. Briefing cards and reviews commonly cite purchase orders, supply confirmations or supplier qualification materials. Data updates per session and per hour, so the validity window of procurement material needs to correspond to the batch scheduled for that session. Retrieval or a form only presents material; authorization status and business actions require separate checking.
For citation tracing, this means the knowledge base search node can associate one or more knowledge bases that share the same vector model, and citation output is an array whose length may be zero. An empty array does not mean the workflow has errored or stopped. The citation limit is measured in tokens. The minimum relevance value applies only during semantic retrieval or when result reranking is used, and it is a filter threshold. Hybrid retrieval merges multiple result paths through RRF, and reranking requires a model and must be enabled according to settings. Retrieval does not perform permission or entity isolation, and a similarity score does not exclude unauthorized records. Knowledge bases using different vector models cannot be associated together; the identifier must be checked first and non-matching bases excluded.
Constraints at the citation tracing step
A hypothetical paper exercise sample is stated here. A brand livestream team drafts one product briefing card line for checking. The material text supplied for comparison has three sentences. First: "Supplier A signed a supply confirmation on 1 March 2025, stating 200 units of Model A in stock, with the quotation valid until 31 March 2025." Second: "This supply confirmation is pending review by the brand procurement lead and has not been signed into effect." Third: "After-sales for Model A is handled by Supplier A, accepted within 7 days from the date of receipt." The draft sentence to be judged is: "The 200 units of Model A in stock have been authorized for livestream, and the supply period covers the whole of April." The comparison object is the three source sentences against the draft sentence. The period is limited to March 2025. The fields are limited to on-hand quantity, review status, after-sales responsible party and quotation validity.
The three source sentences establish a pending review state and a quotation validity end date of 31 March 2025. The draft sentence asserts an authorization and an April supply period. Neither appears in the source text. The after-sales clause is present in the source and is not contradicted by the draft, but the draft does not restate it.
Configuration and implementation checks
The table below uses hypothetical trial starting values tied to this sample. These values are not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the base containing the supply confirmation and the briefing card draft | Check that vector model identifiers match; exclude bases with different models |
| Search mode | Hybrid retrieval | Compare against semantic-only retrieval for hits on model number and subject |
| Citation limit | Hypothetical starting point of 3000 tokens | Compare against the full length of the briefing card draft and observe truncation |
| Minimum relevance | Hypothetical starting point of 0.3 | Compare recall counts against running with the filter off |
| Result reranking | Enabled | Compare ordering against reranking disabled; a rerank model must be available |
| Citation output handoff | Citation array passed to the answering step | Use a paper input of [] to check whether the downstream step continues |
Three common errors
- Treating a pending review state as granted authorization. The sample source states the supply confirmation is pending review and not yet signed into effect. The draft sentence claims authorization for livestream. This claim is unsupported by the source text.
- Treating quotation validity as a supply period. The source states the quotation is valid until 31 March 2025 and does not state any April supply arrangement. The draft sentence claims the supply period covers the whole of April. This is an evidence gap.
- Treating a retrieval hit as a permission confirmation. Citation output length can be zero, and retrieval does not perform authorization judgment. A hit on the supply confirmation does not mean the batch has been approved for livestream.
How to verify the setup
- Input the three source sentences and the draft sentence. Compare the four fields of on-hand quantity, review status, after-sales responsible party and quotation validity. When the draft and source share the same subject, the same period and the same field, and the values are mutually exclusive, record a contradiction. When the source does not state a value, record it as pending verification.
- Input a paper citation array of []. Compare whether the downstream answering step continues to execute. The output path running through without error satisfies the empty-citation handoff condition.
- Input a query containing the Model A number. Compare the hit entries from hybrid retrieval against semantic-only retrieval. When the model number and subject both appear in the cited source text, the tracing check condition is satisfied. When no hit occurs, record the gap and do not assert that recall is guaranteed.