What the documents look like
Dealer ordering events serve offline dealer reservations and do not directly manage platform storefronts. Product reference and selection questions therefore draw on four document families: the ordering handbook, the ordering contract, the sample list, and the credit approval record. These documents are updated by quarter or by year, and a document version corresponds to an ordering batch. When a selection question is answered from an older handbook, the answer can name a model that has stopped being supplied. The brand channel department, regional dealers, and product planning all work from the same materials, so the retrieval scope must first be checked against the selected knowledge bases and their document versions. Only knowledge bases that share the same vector model can be associated together.
The ordering handbook carries the current season's model list, supply cycle, and order deadline. The ordering contract records what a dealer has committed to order. The sample list identifies which physical samples correspond to which models. The credit approval record carries the credit decision for a dealer. These four documents differ in purpose and in the stage they describe, so a citation that is correct for one does not automatically support a statement about another.
Constraints at the citation tracing step
The dataset search node outputs cited content as an array, and the array length can be zero. Even when no content is found, the output path still runs through. Search mode can be semantic search, full-text search, or hybrid search. Hybrid search merges multiple retrieval paths through RRF, and result re-ranking uses a ReRank model to sort candidate text a second time. The re-rank score is 0-1 and can be used for filtering. Minimum relevance takes effect only under semantic search or when result re-ranking is used, and it is a filter threshold. The citation cap controls, in tokens, the maximum amount of content cited per search. Question optimization adds one model call before retrieval and is used for coreference resolution and question expansion. Image retrieval requires a corresponding image input and model capability, and re-ranking requires a model that is enabled according to its setting. These optional paths do not all execute on every request. The retrieval step does not perform permission or entity isolation, so whether a retrieval result is passed to the answering step must be checked item by item.
Hypothetical paper exercise sample: a brand channel department holds an excerpt of the "2026 Spring Ordering Handbook" in the ordering event system. The source text reads: "This season's featured model is the Type A outdoor jacket, with a supply cycle of 45 days after order placement, and an order deadline of March 10, 2026. The Type B shell jacket is a previous-season model and is no longer accepting new orders this season." A draft sentence reads: "The Type B shell jacket can still be ordered this season, with a supply cycle of 30 days." The task is to determine whether this draft sentence is consistent with the handbook source text.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Only the base containing the "2026 Spring Ordering Handbook" | Check the vector model identifier of the selected base; exclude a base whose identifier differs, and do not run cross-model retrieval |
| Search mode | Hybrid search | Model names and dates need full-text recall, and natural-language selection questions need semantic recall; compare with semantic search alone |
| Result re-ranking | Enabled | When candidates are numerous, filter by re-rank score; compare with the ordering result when re-ranking is not enabled |
| Citation cap | Set to 2000 tokens as the starting point for this exercise | The handbook excerpt is short; this value is used to observe citation trimming, compared with a cited content volume of 4000 tokens |
| Minimum relevance | Set to 0.3 as the starting point for this exercise | Used to observe the change in cited item count after filtering, compared with a filtering result of 0.5 |
| Question optimization | Enabled | Selection follow-up questions often contain references; compare with the retrieval terms when it is disabled |
Three common errors
- Treating retrieval output as authorized execution. The search node returns a citation array, which does not mean the ordering contract has been signed, credit has been approved, or an order has been submitted. Check the input as the retrieval result plus the draft sentence, compare the model, supply cycle, and deadline in the draft sentence against the handbook source text, and observe whether only the source text is cited without any business write being triggered.
- Treating minimum relevance as a permission or entity filter. This value is a filter threshold and does not exclude unauthorized records. Check the input as the retrieval results for the same query inside and outside the authorized scope, compare whether the returned citations still contain dealer records without access rights, and observe whether a separate permission-check implementation is needed.
- Mixing knowledge bases with different vector models. Only bases with the same vector model can be associated together. Check the input as the vector model identifiers of two bases, compare whether the identifiers match, and observe whether a mismatched base is excluded before association, without running cross-model retrieval to compare missed recall.
How to verify the setup
- Test with the hypothetical sample source text and draft sentence. The input is the "2026 Spring Ordering Handbook" excerpt and the draft sentence. Compare the model, supply cycle, and deadline fields in the draft sentence against the corresponding fields in the source text, and observe whether the answer cites the source text and states that the Type B model no longer accepts new orders. If the answer does not cite the source text or rewrites it as still orderable, the condition is not met.
- Check downstream handling with a paper empty array. The input is a clearly specified paper retrieval output of
[]. Compare whether the downstream answering node continues to execute with an empty citation, and observe whether a selection conclusion appears without a cited source. This check does not claim that a fictional rule retrieval will necessarily return an empty array. - Test document timeliness with the current version and a historical version of the handbook. The input is the same selection question pointed respectively at the current version and the previous season's version. Compare the version identifier and deadline cited in the answer, and observe whether a version mismatch is recorded as pending verification. Whether a citation is adopted still depends on valid-version review, and no actual test is claimed.
The retrieval step does 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. A pending review does not demonstrate a lack of operating permission, and signing, reaching cooperation, and contract effectiveness are three distinct facts. Where a document field is not provided, it remains unknown, and the missing evidence is recorded as a gap rather than filled with an assumed authorization or effective clause.