What the documents look like
Sample inspection for livestream selling happens before a broadcast. The decision being supported is whether a product can be offered and whether it suits on-camera demonstration. Working materials include supplier files, testing-agency reports, platform negative lists and historical return rates. Day-to-day documents are sample evaluation forms, quality inspection reports, supplier qualifications and prohibited-item lists. In store-service and internal-training work these materials serve two purposes: explaining a selection conclusion to store staff, and letting internal staff check records against one shared standard.
The retrieval step only recalls relevant material and outputs cited content. Citations are returned as an array whose length can be zero. When nothing is retrieved, the output path still completes.
Constraints at the retrieval step
The knowledge-base search node can link one or more libraries that use the same vector model. The smallest search unit is the library; a collection is only a classification layer and does not affect search results. Before retrieval, question optimization can perform coreference resolution and question expansion. Candidates are then recalled through semantic search, full-text search or hybrid search; hybrid search merges multiple result paths through RRF, and result reranking uses a ReRank model to reorder and produce a 0-1 score. The minimum relevance value takes effect only under semantic search or when result reranking is used. The citation limit is measured in tokens. Image retrieval requires matching image input and model capability, and reranking requires an available model enabled in settings. These are optional paths and do not all execute on every request. Retrieval and similarity scores do not perform permission or entity isolation; a recalled result is not an authorized result.
Configuration and implementation checks
The following uses one hypothetical drill sample. Hypothetical drill sample: the store training handover sheet reads, "Sample evaluation form YP-001 submitted by Supplier A, status awaiting quality inspection, submission date 3 March 2025; quality inspection report JZ-001 for the same batch, status issued, covering YP-001." The draft sentence to be judged is, "YP-001 has completed quality inspection and can go on air directly." The source states only that the report was issued. It does not state that the inspection conclusion was a pass, and it does not state broadcast authorization. The draft sentence is unsupported.
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Link only one sample-inspection library with a single vector model | Check the vector model identifier first, exclude different models, and do not compare across models |
| Search mode | Run semantic search and full-text search once each | Use the same question to compare the recalled content of the two modes |
| Citation limit | Start at 800 tokens | Use short and long real materials to observe whether content is truncated, then compare against a larger limit |
| Minimum relevance | Start at 0.3 | Label relevant and irrelevant samples first, then compare filtering against a higher threshold |
| Question optimization | Run once with it off and once with it on | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Run once after enabling it in settings | Compare ordering before and after reranking and confirm the candidate set is sufficient |
The starting values are starting points for this page's hypothetical drill only. They are not official defaults.
Three common errors
- Treating retrieval recall as authorization. Input is the handover sheet text above together with the draft sentence. If the output moves from "issued" to "completed quality inspection, can go on air," the comparison is between the source status field and the draft conclusion. The observable result is that the draft sentence introduces a conclusion the source does not carry.
- Linking libraries across vector models. Input is two libraries with different vector model identifiers. The comparison is between the vector model identifiers of each library. The observable result is that mismatched identifiers are excluded, and no cross-model retrieval is run to compare missed recall.
- Treating optional paths as always executed. Input is a plain-text question with no reranking model configured. The comparison is between this input and the enabled settings. The observable result is that image retrieval and reranking do not participate in this request.
How to verify the setup
- Input the handover sheet text above and the draft sentence. Compare the source status field with the draft conclusion. The observable result is that the draft sentence is judged unsupported, and the two evidence gaps for inspection conclusion and broadcast authorization are recorded.
- Input two libraries with different vector model identifiers. Compare the identifiers. The observable result is that mismatched libraries are excluded, and the actual observation is recorded without inferring runtime behavior.
- Input one plain-text question. Compare this input with the enabled settings. The observable result is that image retrieval and reranking do not participate. Separately, use the explicit paper input
[]to check the downstream design for handling empty citations, without claiming that any rule-based retrieval necessarily returns[].