What the documents look like
Livestream after-sales work clusters after a single session's traffic peak. Refund applications and reshipment orders arrive in dense batches within the same window, and host commitments often remain as spoken statements inside customer-service IM records. Operating reviews and exception retrospectives draw mainly on livestream order lists, refund applications, reshipment orders and script templates. These materials update after each session and daily, so the same order can carry field states from different points in time across different documents.
The knowledge-base search and recall step organizes these documents by library and returns citable reference content for the review. Reference content is output as an array, and its length can be zero. That output link still completes even when no content is found.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the library; a collection serves only to classify and manage data and does not affect search results. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result channels through RRF, and result reranking uses a ReRank model to produce a score between 0 and 1. The reference limit is measured in tokens. The minimum relevance value applies only during semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Retrieval itself does not constitute authorization, and it does not perform entity isolation or permission filtering.
Hypothetical paper exercise: a livestream after-sales team organizes materials after the 18 June 2025 session. Source text one: "Order A1001; the buyer submitted a refund application on 19 June; status is pending review." Source text two: "Order A1001; reshipment order B2001 has been created; status is pending warehouse processing." Source text three, a script template fragment: "This session promises seven-day no-reason returns; specifics follow platform rules." Draft sentence to be judged: "Order A1001 has completed the refund and has been reshipped." This draft sentence conflicts with the pending-review and pending-warehouse-processing states in source texts one and two and is unsupported.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the exercise |
|---|---|---|
| Associated knowledge bases | The node can associate one or more knowledge bases that use the same vector model | First check whether the order-list library and the script library carry the same vector model identifier |
| Search mode | Semantic retrieval, full-text retrieval, hybrid retrieval | Order number A1001 is an exact identifier; full-text or hybrid retrieval is appropriate |
| Reference limit | Unit is tokens; controls the content length referenced per search | This hypothetical trial starting point is 2000 tokens, chosen because order and status fields are short; compare against a smaller value for completeness |
| Minimum relevance | A value from 0 to 1; applies only during semantic retrieval or result reranking | For "refund application status" questions, label relevant and irrelevant samples first, then adjust |
| Question optimization | Coreference resolution and question expansion; adds one model call before retrieval | When a review follow-up asks "was that order refunded," check whether the subject still points to A1001 |
| Result reranking | ReRank model performs a second ordering; score from 0 to 1 | Enable when candidates are numerous; compare against the ordering without it |
Three common errors
- Treating a retrieval hit as confirmation of business status. The input is the draft sentence "Order A1001 has completed the refund and has been reshipped." The comparison objects are the pending-review state in source text one and the pending-warehouse-processing state in source text two. The observable result is that the draft sentence's status field disagrees with both source materials, and it is recorded as unsupported.
- Forcibly associating knowledge bases that use different vector models. The input is the vector model identifier of each library. The comparison is whether the two identifiers are the same. The observable result is that different identifiers mean the libraries cannot be associated together; record the identifiers and exclude the non-matching one.
- Treating pending review as already processed. The input is "status is pending review" from source text one. The comparison is against "has completed the refund" in the draft sentence. The observable result is that pending review does not equal completion, and the evidence gap remains when no approval result is available.
How to verify the setup
- Input source text one, source text two and the draft sentence. Compare the draft sentence's status against the status fields in both source materials. The observable result is that mutually exclusive values mean the draft is unsupported, and the evidence gap is recorded.
- Input order number A1001 and run full-text or hybrid retrieval against the order-list library. Compare whether the returned reference contains that order number and its status field. The observable result is that reference content matching the source material meets the expectation; a miss is recorded as not recalled, without inferring a cause.
- Input an empty array [] as reference content and check whether the downstream design completes with a length of zero. The observable result is that the downstream does not treat the empty reference as having content, which meets the expectation. All of the above are paper checks; none were executed, and no approval, refund or reshipment action occurred.