What the documents look like
Livestream after-sales work clusters after a single broadcast ends, and the promises a host made on air shape most of the tickets that follow. Refund applications, reshipment orders and script templates tend to circle back to the same live session, so drafting copy means locating the original wording tied to that session before writing anything new. The documents in daily use are livestream order lists, refund applications, reshipment orders and script templates. Sources include shop after-sales records, customer service IM, warehouse WMS and logistics tracking. Updates arrive after each session and daily, which means commitment material in a knowledge base needs to be separated by session and date. When retrieval cannot distinguish sessions, wording from one broadcast can be pulled into copy for another.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Bases built on different vector models cannot be associated together. When candidate bases appear, check the vector model identifier first, exclude those that do not match, and record what was observed. Comparing missed recall across models by running a cross-model search is not a valid step.
Search treats the whole base as the smallest search unit. Collections only classify data and do not affect search results. Reference content is returned as an array whose length can be zero, and the output path still completes when nothing is found. Downstream design for empty results can be checked with an explicit paper input of an empty array. A claim that a fictional rule search always returns an empty array cannot be made.
Hypothetical paper exercise: the knowledge base holds a script template from one live session reading, "This session's promise: orders may request reshipment within seven days of signing; reshipment orders must include a logistics exception photo; this promise applies only to orders from this session and not to other sessions." A separate refund application record reads, "Order A-001, signing date assumed 10 March 2024, application date assumed 20 March 2024, application type reshipment." The draft sentence under review is, "This order meets the session promise and reshipment can be arranged." The subject is order A-001, the object is the reshipment promise, the state is an application filed ten days after signing, and the scope is this session. The source promise allows seven days from signing, so the application date falls outside that window. This is a mutually exclusive value in the same field and counts as a contradiction. If the source had not stated a signing date, the result would be unsupported and the evidence gap would be recorded.
Configuration and implementation checks
The table below covers settings and implementation checks tied to this page's documents and primary task. Numeric values are starting points for this hypothetical exercise, with the reason for choosing them and the comparison object named.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only the base holding this session's script template | Check each base's vector model identifier first, exclude different models, then compare whether results fall within this session's material |
| Search mode | Test semantic and full-text search separately | Use the order number and the promise sentence from the sample, compare how numbered items and natural-language items are matched |
| Reference limit | Starting point covers the token volume of the two source passages | Test with short and long material to see whether content stays complete, then adjust the limit |
| Minimum relevance | Starting point is a middle value | Label relevant and irrelevant passages in the sample first, then compare whether irrelevant passages decrease after threshold changes |
| Question optimization | Enabled for follow-ups with references, disabled for explicit questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled when candidates are numerous | Compare the ranking position of the same candidate before and after reranking, confirm model capability and enablement state |
Three common errors
- Treating a semantic recall or relevance score as a permission or entity filter. Retrieval does not grant authorization, and a similarity setting does not exclude unrelated business lines or unauthorized records. Error: assuming a higher minimum relevance keeps other sessions' material out of results. Check: input this session's and another session's material, compare whether returned content still includes the other session, and observe whether filtering acts on relevance alone.
- Describing RRF in hybrid search as a toggle. RRF is a fusion process for hybrid results, not a standalone configuration item. Error: searching the configuration for an RRF switch. Check: review the search mode options, compare actual returns under semantic, full-text and hybrid modes, and confirm fusion happens inside hybrid mode.
- Treating a pending or in-progress status as granted authorization or completed action. A pending review does not show that an operation is unauthorized, and signing, partnership agreement and contract effectiveness are three separate facts. Error: writing "reshipment can be arranged" from a submitted reshipment application. Check: compare the application record with the promise source, confirm whether the signing date and application date fall inside the promise window, record a contradiction when outside, and record an open item when the date is absent.
How to verify the setup
- Input the promise source and the draft sentence from the sample, then compare the draft's application time against the seven-day window in the source. If the application date falls beyond seven days after signing, the contradiction condition is met and the draft sentence should be judged unsupported. If the source does not state a signing date, record the result as unsupported and keep the evidence gap open.
- Input order number A-001 from the sample and run semantic and full-text search separately, then compare the candidates each mode returns. If full-text search locates the source containing the order number while semantic search returns generalized content, record that difference for deciding mode choice on numbered items.
- Input the paper empty array and check whether the downstream design continues through the output path. If downstream proceeds under the empty array, the empty-result continuation condition is met. A claim that a fictional rule search always returns an empty array cannot be made; only the observation under this paper input is recorded.