What the documents look like
Livestream after-sales work clusters after a single session's traffic peak, and refund requests are strongly tied to what the host promised on air. The working materials are livestream order lists, refund applications, replacement orders and script templates. In cross-border operations these documents often mix Chinese and English, so an agent handling one commitment may need to locate the original wording of that commitment across more than one knowledge base.
The retrieval layer described here is the knowledge base search node. It can associate one or more knowledge bases that share the same vector model, and it outputs references as an array whose length may be zero. A zero-length output still lets the downstream path run. Retrieval is recall, not authorization. A similarity score does not separate business entities or permission scopes, and a relevance threshold is not an entity or permission filter.
Tracing a citation means comparing an answer sentence against the retrieved source text item by item, and treating the effective version review as the authority. The following paper exercise is a hypothetical sample, not a record of any real order, customer or approval.
Hypothetical paper sample: a livestream session dated 12 March 2025, order number SAMPLE-0312-07, where the buyer files a refund request concerning replacement timing. Knowledge base A (Chinese script templates, vector model M1) states: "Replacement orders are arranged within 48 hours after the warehouse confirms an error or omission." Knowledge base B (English refund policy, vector model M1) states: "Replacement is arranged within 48 hours after warehouse confirmation." Draft sentence under review: "This order has received replacement approval and will ship within 48 hours." Handover request: "Search knowledge bases A and B, confirm the replacement timing wording and provide the citation source."
Constraints at the citation tracing step
The knowledge base search node associates knowledge bases that use the same vector model. Knowledge base A and knowledge base B both use M1, so they can be associated together; a library using only M2 cannot join that association. The node returns a reference array, and an empty array is a valid output rather than a signal that the workflow failed.
Search mode, reference cap and minimum relevance are settings on the search side. The reference cap is expressed in tokens per search, not as a top-k item count. Minimum relevance is a 0-1 value that filters low-relevance results, and it takes effect only in semantic search or when result reranking is used. It is a filtering threshold, not a final ordering rule, and it does not exclude unrelated businesses or unauthorized records.
Result reranking uses a ReRank model to reorder candidates and produces a 0-1 score. Hybrid search combines semantic and full-text recall through RRF, which is a fusion step inside the retrieval pipeline rather than a switch to turn on. Question optimization adds a model call before retrieval and can help multi-turn or short questions; it also adds latency.
The tracing step itself is a manual comparison. The retrieved array must be handed to the answering stage, and each answer sentence must be checked against the cited source text. Whether a citation is adopted still depends on the effective version review. The sample's handover request only asks for a search; it does not show that a search ran, that citations were verified, or that any replacement was approved.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Knowledge base A and knowledge base B | Both use vector model M1, which is required for joint association; a library using only M2 is excluded by identifier comparison |
| Search mode | Hybrid search | The sample contains both the order number SAMPLE-0312-07 and a natural-language question, so semantic and full-text recall are combined through RRF |
| Reference cap | 2000 tokens as a hypothetical starting point | The sample material is short; comparing against 500 tokens shows whether the English paragraph is truncated. This is not an official default |
| Minimum relevance | 0.3 as a hypothetical starting point | Comparing against 0.5 shows how the number of recalled items changes; the value applies only in semantic search or with reranking |
| Result reranking | ReRank enabled | With several candidates, a second ordering produces a 0-1 score; comparing with reranking disabled shows ordering stability |
| Reference handoff | Output array passed to the answering stage | Each answer sentence is compared against the cited source text; an empty array follows the zero-length downstream path |
Three common errors
- Treating a retrieval hit as replacement approval. The sample knowledge bases state timing wording only. They contain no approval record, so the draft sentence "has received replacement approval" lacks supporting evidence.
- Treating minimum relevance as a permission or entity filter. The similarity threshold filters low-relevance results only. It does not exclude unrelated businesses or unauthorized records, and retrieval itself is not authorization.
- Treating a pending handover request as an executed action. The request asks for a search and a citation source. It does not show that retrieval finished or that any citation was verified, so tracing cannot be called complete on that basis.
How to verify the setup
- Input the sample draft sentence together with the source text of knowledge bases A and B, then compare whether "has received replacement approval" matches the sources on the same subject, the same period and the same field. The sources contain timing wording only and state no approval status, so the sentence remains unsupported.
- Input order number SAMPLE-0312-07 and compare whether the returned references point to source text in knowledge base A or knowledge base B. If the array is empty, check the downstream handoff design against a zero-length output; do not infer that no result must occur.
- Input a request to associate a library that uses only M2, then compare the vector model identifier of the selected knowledge bases. When M2 differs from M1, exclude that library and record the actual observation; do not run a cross-model retrieval to compare missed recall.
The optional paths in the overall retrieval 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. A knowledge base using a different vector model cannot be associated, and no runtime outcome such as missed recall should be asserted for that case; check the identifier, exclude the mismatched library and record what is observed.