What the documents look like
Livestream violation appeals respond to penalties that have already been issued, so the work involves reconstructing an evidence chain rather than preventing a future event. The materials come from the platform violation center, livestream replays, customer service records, and the product qualification library. The documents handled day to day include violation notices, livestream recordings, product qualification files, and appeal statements. These materials are updated irregularly, with each violation triggering a round of organization.
The task on this page is product and marketing copy drafting. The core action is to locate, inside existing materials, the source text that could support a proposed sentence, then judge whether that sentence is actually supported. Knowledge base retrieval and recall supplies the reference content drawn from violation notices, recording records, and qualification files so that each drafted sentence can be compared line by line. Whether the material is complete, whether the subject is clear, and whether a status falls inside its valid range directly determine whether a proposed sentence can stand.
Three differences shape this work. First, the material formats diverge sharply: violation notices lean toward identifiers and line items, recording records lean toward timestamps and scene descriptions, and qualification files lean toward subject and validity period. Second, retrieval here is not authorization: reference content only tells the drafter what the documents say, and a similarity score is a relevance number rather than a business ownership or permission judgment. Third, a missing field is not a denial. When a qualification file states a subject and a validity period but does not address a particular claim, that silence supports neither the sentence nor a contradiction of it.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. During setup, verify the vector model identifier of each candidate library and exclude any that does not match, recording what was actually observed rather than assuming a runtime outcome.
Search mode selection depends on the material. Semantic retrieval relies on vector distance and suits natural-language questions and semantically close content. Full-text retrieval relies on keyword hits and suits identifiers, model numbers, proper nouns, and error codes. Hybrid retrieval uses both and merges the two result sets through RRF, which is a fusion step inside hybrid retrieval rather than a separate switch. ReRank performs a second ordering pass over candidate text and produces a 0-1 score. The reference cap is measured in tokens, and the minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call before the actual retrieval.
Two limits deserve care. The minimum relevance value is a filter threshold, not a final ordering rule; once question optimization, hybrid retrieval, image retrieval, or reranking is active, results may be merged across multiple recall paths and need not follow a single vector similarity order. Optional paths shown in the overall retrieval diagram do not all execute on every request: image retrieval requires a corresponding image input and the relevant model capability, and reranking requires a rerank model enabled in the settings.
Hypothetical paper exercise sample: the source material reads, "Store A received a platform violation notice on 10 March 2025, numbered T-20250310-07, stating that a promotional phrase for product X during the 8 March livestream involved a claim about efficacy. The qualification file for product X states the subject as Store A, valid until 31 December 2025. The recording record shows that at 20:15 on 8 March the host's spoken content matched the description in the notice." The proposed sentence to be judged is, "Store A's livestream on 8 March 2025 described the efficacy of product X within the scope of the qualification file." The comparison holds subject, date, and field constant: the notice finds a problem with the efficacy claim, while the qualification file states only the subject and the validity period and does not state that the efficacy claim is compliant. The proposed sentence is not supported, and the evidence gap must be preserved.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Violation notice library and qualification library, with vector model identifiers verified as matching | Association requires matching identifiers; check each one and exclude any that differ |
| Search mode | Run semantic retrieval and full-text retrieval once each | Compare the content recalled by the two paths on the same test question, then decide whether to use hybrid retrieval |
| Reference cap | Hypothetical starting value of 3000 tokens | Test whether reference content stays complete using materials of different lengths, then adjust after comparing with a shorter starting value |
| Minimum relevance | Hypothetical starting value of 0.4 | Label relevant and irrelevant samples first, then compare what remains after filtering before adjusting the threshold |
| Question optimization | Off first, with one retrieval using the original question | Turn it on, then compare whether the optimized question has altered the subject or object |
| Result reranking | Not enabled initially | When candidates are numerous and the text question is clear, enable it and compare the rerank score against the filtering effect |
Three common errors
- Treating a retrieval score as an authorization or ownership judgment. The input is the proposed sentence together with the recalled reference content; the comparison is whether the source text states the same subject, the same period, and the same field; the observable result is a similarity number. A high score does not mean the material supports the sentence, and line-by-line comparison with the source text is still required.
- Treating a missing field as a denial. The input is the qualification file text; the comparison is whether the file states a field about efficacy claim compliance; the observable result is that the field is absent. An absent field can only be recorded as unsupported or pending verification, never as a conflict or a denial.
- Forcing an association across knowledge bases with different vector models. The input is the vector model identifier of two knowledge bases; the comparison is whether the identifiers match; the observable result is that they do not. Exclude the non-matching library and record the observation; do not run a cross-model retrieval to compare missed recalls.
How to verify the setup
- Check the notice text against the proposed sentence using the hypothetical sample. The input is the source text for notice number T-20250310-07 together with the proposed sentence; the comparison covers the same subject Store A, the same date 8 March 2025, and the same field of efficacy claim; the observable result is that the notice finds a problem while the qualification file does not state compliance. The condition is met when the proposed sentence is marked as unsupported.
- Check how an empty result is carried downstream. The input is a paper input of an empty reference array; the comparison is whether the downstream drafting step continues through the zero-length reference; the observable result is that the chain is not interrupted. The condition is met when the downstream step accepts the empty reference without breaking.
- Check the subject before and after question optimization. The input is the original question and the optimized question; the comparison is whether the three elements of Store A, product X, and 8 March 2025 are retained; the observable result is that the optimized question still contains the same subject and object. The condition is met when no element has been substituted or added.