What the documents look like
Brand dealer ordering events serve offline dealer reservations and do not directly manage platform storefronts. The working data comes from the ordering event system, historical sales, dealer inventory, and financial credit approval. Everyday documents include the ordering handbook, the ordering contract, the sample list, and credit approval records. Update frequency is quarterly and yearly.
In order fulfillment and after-sales Q&A, dealers often raise follow-up questions about new product policy, order quantities, and credit lines. Answers need to cite the original text of the ordering handbook, contract, or credit approval, and mark the source and its valid scope. Material versions update by quarter, and handbooks and contract terms can differ between quarters, so citation work must check the material version and the applicable period.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model for vector search. Knowledge bases built on different vector models cannot be associated together; the identifier must be checked first and non-matching bases excluded. Search output is an array of citations whose length can be zero, and even when no content is found the output path still runs through. An empty array does not mean the workflow has errored and stopped.
Citation content must be handed to the answering step, and each answer must be checked against the cited original text item by item. Source display is checked with traceable samples; thinking process or answer fluency does not substitute for source evidence. Material timeliness is checked by testing sources with the current version and a historical version, and whether a citation is adopted still depends on review of the valid version. Retrieval and similarity scores do not carry permission or entity isolation, and the same-vector-model requirement for associated knowledge bases does not extend to business registration sites or other external systems.
Hypothetical drill sample: material original text — "2025 Autumn Ordering Handbook, Chapter 3: For the A-type product listed in this ordering event's sample list, dealers may submit reservation quantities from September 1 to September 15. Credit Approval Form No. X-2025-018 states that this dealer's credit line is RMB 500,000, valid until December 31, 2025." Draft sentence to be judged: "This dealer's reservable amount for this ordering event is RMB 500,000, and the quotation is valid until December 31." The draft sentence conflates the credit line with quotation validity and does not check the sample list or contract terms.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Checking method |
|---|---|---|
| Associated knowledge bases | Select the ordering handbook base and the contract base that share one vector model | Check base identifiers and vector models for consistency, exclude bases with different models |
| Search mode | Hybrid search | Uses semantic recall and full-text recall together, merged through RRF |
| Citation limit | Hypothetical starting point 2,000 tokens | Compare with the actual length of the sample original text to confirm citation content is complete |
| Minimum relevance | Hypothetical starting point 0.5 | Compare with the rerank score to filter low-relevance results |
| Result rerank | ReRank enabled | Check whether the rerank score is used for filtering |
| Question optimization | Enabled for consecutive follow-up questions | Check whether the completed result matches the current question |
The citation limit uses tokens as its unit, not top k. The reason for the hypothetical 2,000-token starting point is that the sample original text runs to a few hundred characters, leaving margin; the decision to adjust is made after comparing with the actual length of the sample original text. The 0.5 minimum relevance is a hypothetical trial starting point for this page, judged against the rerank score for filtering. Thinking budget fields govern thinking intensity and are not used for the citation token limit.
Three common errors
- Treating the credit line as the orderable amount. The draft sentence says "reservable amount is RMB 500,000," while the material original text says the credit line is RMB 500,000; these are different fields. The check compares the field name and value in the draft sentence against the original text, and marks the draft as unsupported when the fields differ.
- Treating quotation validity as the supply period. The material original text states the credit is valid until December 31, while the draft sentence says "the quotation is valid until December 31," with a different subject and field. The check compares the object the validity period belongs to, and records an evidence gap when the objects do not match.
- Associating knowledge bases across vector models. If the ordering handbook base and the contract base use different vector models and are associated together, retrieval cannot run normally. The check verifies base identifiers and vector models first, excludes non-matching bases, and then runs retrieval.
How to verify the setup
- Input the hypothetical sample original text and the draft sentence, and compare whether the draft sentence's field shares the same subject, the same period, and the same field as the original text. If the fields differ, the output is unsupported; if the fields match and the values are mutually exclusive, the output is a contradiction.
- Input an empty array [] to check the downstream design, and observe whether the answering step accepts an empty citation. If the downstream design does not handle an empty array, record it as pending verification and do not claim that a fictional rule retrieval necessarily returns [].
- Input the ordering handbook base and the contract base that share one vector model, and observe whether retrieval returns a citation array. If the returned array length is greater than 0, check whether the cited original text and the answer correspond item by item; if the returned array length is 0, record the actual observation and do not infer a cause of error.