What the documents look like
A brand-run livestream room carries its own scripts, inventory and after-sales obligations, so the working documents are the livestream script, the product explanation card, the shift schedule and the post-session review. In cross-border operation, one product can have a Chinese explanation card and English or Southeast Asian language materials side by side, and retrieval has to match across those languages. Scripts and explanation cards are refreshed per session; schedules and reviews accumulate day by day. Once retrieval settings are fixed, the same knowledge base is reused across sessions.
For this page the relevant consequence is that a single product fact may exist in several languages at once, and the retrieval step is the only place where a Chinese question can reach an English knowledge point. The script and the explanation card are the materials that carry the sellable quantity, price and offer window, so retrieval accuracy on those fields matters more than on general commentary.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; the vector model identifier is checked first and non-matching bases are excluded, without running a cross-model retrieval to compare what would be missed.
Semantic retrieval works through vector distance and supports cross-language understanding, so a Chinese question can match an English knowledge point, but precision depends on how well the model was trained. Full-text retrieval suits numbers, model designations and proper nouns. Hybrid retrieval combines semantic and full-text results through RRF; RRF is a fusion step in the pipeline, not a switch that can be turned on. The minimum relevance value is a filter threshold and takes effect only under semantic retrieval or when result reranking is used; it does not perform permission or entity isolation. The reference limit is measured in tokens, and the output reference array can have length zero, so the downstream path still runs when nothing is found. Question optimization performs coreference resolution and question expansion before retrieval, and it adds a model call, so it raises latency.
A hypothetical paper exercise sample, used only for checking: the material text to be checked reads, "Session of 12 April 2025, Model A vacuum flask Chinese explanation card states: sellable stock for this session 800 units, price 79 yuan, offer closes 15 April 23:59. The corresponding English explanation card states: Stock 800 units, price 79 CNY, offer ends Apr 15 23:59." The draft sentence to be judged reads, "The Model A vacuum flask English explanation card shows the offer is valid long term and orders can be placed at any time." The judgement requires the same subject, the same period and the same field. The source states a closing time of 15 April 23:59; the draft claims long-term validity. These are mutually exclusive values in the same field, so the draft contradicts the source. If the material did not state a closing time, the claim would be recorded as unsupported, not as a denial.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Check that the vector model identifiers match before associating bases together | The Chinese and English explanation cards must share a vector model to be searched in one base |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | Use the sample Chinese question to observe whether the English explanation card is reached |
| Reference limit | Unit is tokens; test completeness with materials of different lengths | A hypothetical starting value of 2000 tokens is chosen for this exercise because one explanation card runs to a few hundred characters and several references must fit; compare against 1000 tokens to see whether content is truncated |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking; label relevant and irrelevant samples before adjusting the threshold | Use the sample draft sentence to observe whether a low threshold pulls in unrelated fragments |
| Question optimization | Performs coreference resolution and question expansion; compare the original question with the optimized question | Check whether the subject is altered by optimization |
The database internals described in the official material, such as the vector index and the storage split between engines, are implementation facts of the platform and are not per-page configuration items.
Three common errors
- Associating knowledge bases that use different vector models. The input is the vector model identifier of each base; the comparison is whether the identifiers are equal; when they differ, the base is excluded. No runtime outcome of a combined association is invented.
- Treating minimum relevance as a permission or entity filter. The input is an unrelated business question; the comparison is whether the retrieved references contain that question. A relevance score does not exclude unrelated records, and retrieval is not authorization.
- Treating a handover request awaiting review as approved and executed. The input is a shift handover request marked as pending review; the comparison is whether a signature or approval record exists. A pending state alone is recorded as unsupported.
How to verify the setup
- Input the sample Chinese question together with the English explanation card text. Compare the retrieved reference against the offer closing time field. The condition is met when that field appears and its value matches the source text.
- Input the sample draft sentence together with the explanation card text. Compare the offer validity field. The draft states long-term validity while the source states a closing time, so the judgement is a contradiction and this is an expected failure branch.
- Input an empty reference array [] and check the downstream design. Compare whether the downstream path handles the empty result branch. No claim is made that a retrieval under invented rules necessarily returns []. If the material lacks the offer closing time field, the evidence gap is recorded and no clause is supplied.
Across these checks, the material needed for the judgement must sit inside the target retrieval scope to which access is held, or be supplied as a clearly labelled manual input. Where it is absent, the gap is recorded rather than filled. Semantic recall and relevance do not carry permission or entity isolation, and the same-vector-model requirement for associated bases does not extend to any external system.