What the documents look like
Livestream product selection and sample quality inspection happen before a broadcast begins. The central question is whether a product can be sold and whether it suits on-camera demonstration. Four document types carry that judgment: sample evaluation forms, quality inspection reports, supplier qualification files, and prohibited-item lists. In order fulfillment and after-sales answering, retrieval over these four types supports return decisions, qualification tracing, and explanations of after-sales risk.
Materials update weekly and per item. The same product can therefore have materials that change by batch. Any retrieval used for an after-sales answer must check which batch and which date range the material covers. A hit from an older batch does not describe the batch that was sold.
A second property matters for this task. Retrieval returns reference content. It does not return an authorization decision. A supplier qualification file appearing in results shows the file was retrieved. It does not show the supplier passed review, that a contract took effect, or that a refund was approved.
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 using different vector models cannot be associated together. Implementation checks the vector model identifier first, excludes non-matching bases, and records the actual observation. Cross-model retrieval is not executed to compare missed recalls.
Search modes include semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process inside hybrid retrieval, not a switch. Minimum relevance applies only in semantic retrieval or when result reranking is used. It is a filtering threshold. It does not perform permission filtering or entity isolation. The reference limit is measured in tokens. Reference output can have length zero, and the output path still completes when nothing is retrieved.
Hypothetical paper exercise sample: the material text reads, "The sample evaluation form submitted by Supplier A (number YP-2024-031, evaluation date March 12, 2024) states: this batch of facial tissues has been sent for testing, and the quality inspection report is pending; the supplier qualification file is within its validity period; the product is not on the platform prohibited-item list." The draft sentence to be judged reads, "This product has completed quality inspection and can go on sale in the livestream." The comparison checks whether the claimed "has completed quality inspection" and the source "quality inspection report is pending" are mutually exclusive values in the same field. The source states the report is pending and does not state the report has been issued. The draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate one base that uses the same vector model | Check the vector model identifier first, exclude non-matching bases, and do not run cross-model retrieval |
| Search mode | Start with full-text retrieval for number YP-2024-031 | Number queries depend on keyword hits; compare hit differences against semantic retrieval |
| Reference limit | Start at a token amount covering the full text of one sample evaluation form | Test with materials of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set a threshold | This value applies only in semantic retrieval or result reranking, and is not used for permission filtering |
| Question optimization | Enable for multi-turn follow-up scenarios, disable for clear single-turn questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Use when a rerank model is available and enabled per settings | Rerank scores can be used for filtering; when not enabled, no such score is produced |
Three common errors
- Treating a retrieval hit as an authorization action. Retrieval returns reference content only. Business verification, approval, database writes, and external synchronization each need their own downstream implementation. Writing "the supplier qualification file is within its validity period" as "the supplier has passed review" treats a retrieval result as an authorization conclusion.
- Writing a pending report as completed inspection. In the hypothetical sample the source reads "quality inspection report is pending," while the draft sentence reads "has completed quality inspection." These are mutually exclusive values in the same field and form a contradiction. When the source does not state the report has been issued, the item is recorded as pending verification.
- Forcing knowledge bases with different vector models into one association. The documentation states the node can associate one or more knowledge bases using the same vector model, and cross-model bases cannot be associated together. Implementation checks the identifier, excludes non-matching bases, and does not fabricate runtime results of a cross-model association.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence "This product has completed quality inspection and can go on sale in the livestream." Compare whether "has completed quality inspection" and the source "quality inspection report is pending" are mutually exclusive in the same field. If the judgment is not supported, the checking logic works. If the judgment is supported, the checking logic is wrong.
- Input number YP-2024-031 and run full-text retrieval and semantic retrieval separately. Compare whether the reference content returned by each path contains the material holding that number. If full-text retrieval hits and semantic retrieval does not, record that difference. If neither path hits, record the gap and do not claim retrieval necessarily returns empty.
- Input a query containing text only and no image. Compare whether returned references come only from the text path. If image description or image vector results appear, the image retrieval path was triggered unexpectedly. If they do not appear, record the actual observation and do not infer that image retrieval is disabled.
When the paper input is an empty array [], the downstream design should accept a reference output of length zero. Do not claim that a fictional rule search necessarily returns [].