What the documents look like
A brand-run livestream team writes its own scripts, holds its own inventory, and handles its own after-sales work, so product materials and selection answers directly shape what hosts say on camera and how the comment section is answered. The working documents are livestream scripts, product explanation cards, shift schedules, and livestream reviews. Reference data comes from Douyin Shop Compass, Juliang Qianchuan, livestream comments, and transaction details. Materials refresh at a per-session and hourly level, so the retrieval layer sees the current version of a script or explanation card, and version differences are handled by manual cross-checking. A typical selection question asks whether a given product may be explained in a specified session, which explanation card applies, and which specifications remain on sale. Retrieval results need to point to a specific document so the floor controller and host can review 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 with different vector models cannot be associated together. Retrieval treats the knowledge base as the smallest search unit, while a collection only classifies data and does not affect search results. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, hybrid retrieval merges both paths through RRF, and result reranking uses a ReRank model for a second ordering. The reference limit is measured in tokens. 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. Reference content is output as an array and its length can be 0. The retrieval step itself does not perform permission or entity isolation and does not decide business authorization; form value collection, approval, database writes, and external synchronization each need their own downstream implementation.
Hypothetical drill sample: the source material reads, "March 14, 2026, session A livestream, explanation card number JK-0314-A, applicable product is Jia-brand insulated cup 500 ml, status is reviewed and usable." The draft sentence to be judged reads, "The Jia-brand insulated cup 500 ml may be explained in the March 14, 2026 session A livestream, based on JK-0314-A." The check compares the subject, object, session, date, and status fields against the source material.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the single vector-model knowledge base holding livestream scripts and explanation cards | Compare against the identifier of a different vector-model knowledge base, exclude the mismatch, and associate the qualifying one alone |
| Search mode | Run the same test questions through semantic retrieval, full-text retrieval, and hybrid retrieval separately | Compare hit behavior for number-type questions against natural-language questions |
| Reference limit | Set 1200 tokens as a starting point | Compare explanation cards of different lengths to see whether returned content is truncated |
| Minimum relevance | Set 0.3 as a starting point | Compare filtering behavior against labeled relevant and irrelevant samples |
| Question optimization | Enable for multi-turn follow-ups and short questions | Compare the original question and the optimized question to see whether the subject was altered |
| Result reranking | Enable when candidates are numerous and the question is clear | Compare ordering of the same candidate batch before and after reranking |
These values are hypothetical trial starting points for this page and do not come from official defaults.
Three common errors
- Treating a retrieval hit as authorization. In the hypothetical drill sample, JK-0314-A has status reviewed and usable, and retrieval can return that explanation card, which only shows the material can be found in the knowledge base. If an explanation card has status pending review, retrieval may still return it, and pending review does not constitute usable authorization. Executable check: input the explanation card number, compare the status field in the returned content against the status field in the source material, and observe whether the status field is missing or differs from the source.
- Forcing association across vector-model knowledge bases. Executable check: input the vector model identifiers of two knowledge bases, compare whether the two identifiers match, and observe whether identifiers that differ were still placed in the same association list; when identifiers differ, exclude the base and do not run a cross-model retrieval comparison.
- Treating minimum relevance as a permission or entity filter. Executable check: input a query that is semantically close to the question but unrelated to authorization, compare the returned content against permission fields, and observe whether permission fields were left unfiltered at the retrieval layer; the retrieval layer does not perform that filtering, so record the observation and route it to the downstream implementation.
How to verify the setup
- For the hypothetical drill sample, input the draft sentence and compare the subject, object, session, date, and status fields in the returned reference against the source material. When all five fields match and the status is reviewed and usable, the draft sentence is treated as supported by the material. If any field is not stated, record it as unsupported or pending verification.
- For a number-type query, input JK-0314-A and compare the returned content from full-text retrieval against semantic retrieval, observing whether full-text retrieval hits the data containing that number. If it does not hit, record that as an observation and do not infer a failure cause.
- For empty-result handling, construct a paper input of an empty array and check whether the downstream design still carries the chain through when the reference length is 0. This check is a paper exercise and does not claim that retrieval under a fictional rule will necessarily return an empty array.