What the documents look like
Livestream product selection and sample quality inspection happen before a broadcast. The work answers whether a product can be sold and whether it suits on-camera presentation. The main sources are supplier materials, testing-agency reports, platform negative lists, and historical return rates. Everyday documents include sample evaluation forms, quality inspection reports, supplier qualifications, and prohibited-sale lists. Materials update weekly and per product. At the retrieval step, several documents about the same product are in scope, and their subjects, periods, and fields do not align. The knowledge base search node can associate one or more knowledge bases that use the same vector model. That restriction directly affects whether selection materials can sit inside a single retrieval scope.
Constraints at the retrieval step
One retrieval combines semantic search, full-text search, image retrieval, question optimization, and reranking according to the input content and search parameters, then merges the multiple result paths through RRF, applies similarity filtering and a reference cap, and outputs reference content. Reference content is output as an array whose length can be 0, which means the output path still completes when nothing is found. The minimum relevance value takes effect only in semantic search or when result reranking is used. Full-text search suits precise lookups such as numbers, models, and proper nouns. Image retrieval requires corresponding image input and model capability; reranking requires a rerank model and must be enabled by configuration. The paths listed in the documented flow do not mean every request executes all of them. Retrieval and relevance scores do not carry permission or entity isolation functions.
Hypothetical paper exercise: a sample evaluation form states "Supplier A, product is a folding electric cooker, sample submission date 10 March 2025, sample quantity 2 units, evaluation conclusion field left blank." A supplier qualification document states "Supplier A, business license valid until 1 May 2026." A pending handover request reads "Please confirm, based on the above materials, that Supplier A's product can go on air." This request exceeds what the materials can support: the evaluation conclusion is missing, and the qualification document states only the entity and validity period, with no on-air authorization recorded.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only selection material bases using the same vector model | Check the vector model identifier of each base; exclude those whose identifiers differ |
| Search mode | First run the same question separately under semantic search and full-text search | Compare returned content under the two modes, then decide whether to use hybrid search |
| Reference cap | Hypothetical starting point of 3000 tokens | Use actual materials of differing length to observe whether content is truncated, compared with results at a longer cap |
| Minimum relevance | Hypothetical starting point of 0.4 | First label relevant and irrelevant samples, then compare filtering results at different thresholds |
| Question optimization | Enable for follow-up questions containing references; disable for explicit questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when a rerank model is available and candidates are numerous | Compare the ranking positions of the same candidate set before and after enabling |
The values above are hypothetical trial starting points for this exercise and are not official default values.
Three common errors
- Treating a retrieval hit as an authorization conclusion. Input "Supplier A's product can go on air," compare the returned sample evaluation form and qualification document, and observe that the evaluation conclusion field is empty and the qualification document records no on-air authorization. The output can only be recorded as pending verification.
- Forcing knowledge bases with different vector models into a joint association. Input the vector model identifiers of the two bases, compare whether they match, and exclude them when they do not, recording only the actual observation without inferring missed recall or redundant results.
- Treating minimum relevance as a permission or entity filter. Input a question unrelated to product selection, compare the returned content after filtering, and note that low relevance affects only whether content is filtered and cannot be used to judge whether a record falls within an allowed access scope.
How to verify the setup
- Input the sample evaluation form text and the qualification document text above, and ask "Can Supplier A's product go on air?" Compare whether the returned references cover both materials and whether a notice of the missing evaluation conclusion appears. The condition is met when the reference content comes from those two materials and no on-air conclusion is generated.
- Input a follow-up question containing a reference, such as "What about its qualification?" Compare the retrieval terms before and after question optimization, and check whether the subject still points to Supplier A. The condition is met when the optimized question retains the original subject and does not substitute another supplier or product.
- Input a precise number query, such as looking up a quality inspection report by model. Compare the returned entries under semantic search and full-text search, and check whether the number is matched in full. The condition is met when the content returned by each mode is recorded, and a miss retains an empty result without claiming recall is guaranteed.