What the documents look like
Dealer ordering events serve offline dealer reservations and do not directly manage platform storefronts. The working materials are ordering handbooks, ordering contracts, sample lists, and credit approval forms. Product information and selection questions draw on these materials to answer model numbers, minimum order quantities, policy terms, and contract clauses. Supporting data comes from the ordering event system, historical sales, dealer inventory, and financial credit records. These materials update by quarter or by year, so a retrieval configuration aligns with the current season's material scope. Identifiers, model numbers, and proper nouns in the materials suit full-text retrieval; natural-language descriptions suit semantic retrieval; hybrid retrieval merges both paths through RRF.
A retrieval step returns recalled reference content. It does not authorize an order, approve credit, or confirm a contract. Forms collect values; business verification, approval, database writes, and external synchronization require their own downstream implementation.
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 built on different vector models cannot be associated together; the vector model identifier of each candidate base is checked first and non-matching bases are excluded. The output is an array of reference content whose length can be zero, meaning the output path still runs when nothing is found. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. Retrieval is recall, not authorization.
Image retrieval requires an image input and matching model capability, and reranking requires an available model enabled in settings. The overall retrieval flow lists optional paths; not every request executes all of them. RRF is a fusion process for combining result paths, not a standalone toggle.
Hypothetical paper exercise sample. The material text supplied for checking reads: "2025 Spring Ordering Handbook, Chapter 3: Sample Type A minimum order quantity is 50 units, order deadline is March 15." A second line reads: "Sample Type A credit approval form status is pending review." The draft sentence to be judged reads: "Sample Type A has received credit approval and can execute the 50-unit minimum order." Compared against the status field in the source text, this draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the ordering handbook base that uses the same vector model | Check the vector model identifier and exclude bases using a different model |
| Search mode | Hybrid retrieval | Compare recall for model-number questions and natural-language questions separately |
| Reference limit | Measured in tokens, starting point 2000 | Test with materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Starting point 0.5 | Label relevant and irrelevant samples first, then compare filtering results as the threshold changes |
| Question optimization | Enabled for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled according to settings | When candidates are numerous, compare ordering before and after reranking and confirm the model is available |
The reference limit is measured in tokens and does not use top k. Question optimization adds one model call before formal retrieval, which improves multi-turn retrieval and increases elapsed time. RRF is the fusion step for hybrid results and is not an independent switch. The minimum relevance value filters low-relevance results and does not replace permission checks.
Three common errors
- Treating retrieved content as an authorization basis. The input is "Sample Type A credit approval form status is pending review." If the output states the request is approved, the pending status has been treated as approval. Compare the approval form status field with the draft sentence and observe whether the two agree.
- Associating knowledge bases that use different vector models. The input is two bases with different vector model identifiers. Associating them directly ignores the same-model requirement. Check each base's vector model identifier, exclude non-matching bases, and record what is observed.
- Treating minimum relevance as a permission filter. The input is an unrelated business question. Expecting this setting to exclude unauthorized records confuses filtering with authorization. Compare reference content before and after a threshold change and confirm that permission remains a downstream system decision.
How to verify the setup
- For the draft sentence "Sample Type A has received credit approval and can execute the 50-unit minimum order," supply a retrieval scope that includes the material text above, then compare the draft sentence with the credit approval form status field. The source status is pending review and the draft sentence states approval; the two values are mutually exclusive in the same field, so the draft sentence is unsupported.
- For the question "What is the minimum order quantity for Sample Type A," supply the associated base and compare whether the reference content contains the minimum order quantity statement from Chapter 3 of the handbook. If the reference content contains that statement with Sample Type A as the subject, recall is effective; if the reference array is empty, record the gap and do not infer that the rule necessarily returns empty.
- For the question "What is the order deadline for Sample Type A," supply hybrid retrieval and compare the reference content after semantic and full-text results are merged through RRF. If the reference contains March 15 and matches the handbook, the check passes; if the two paths disagree, record the difference and do not assume the first result is the correct hit.