What the documents look like
Ordering events serve offline dealer reservations and do not directly manage platform storefronts. In this category, the reference sources used for store service and internal training come from four material types: the ordering handbook, the ordering contract, the sample list, and the credit approval form. These materials update by quarter or by year. Explaining new product policy, collecting order quantities, and approving credit lines are the main time-consuming steps.
Citation tracing here has a direct purpose: store service scripts and internal training content must return to the correct version of the original material. A script must not relay policy to a dealer on the basis of an outdated handbook or a contract that has not taken effect.
The knowledge base search node outputs a citation content array whose length can be zero. An empty array means the output path completes; it does not mean the workflow reported an error and stopped. Linked knowledge bases must use the same vector model, and bases with different vector models cannot be linked together. The minimum relevance score is a filter threshold that takes effect only in semantic retrieval or when result reranking is used; it does not perform permission or entity isolation. Retrieval itself is not authorization. Whether a credit approval passed, whether a contract took effect, and whether a termination completed must all be judged from the signing and approval status of the corresponding material. A retrieval hit or a filled-in form cannot substitute for that judgment.
Assumed exercise sample: page 3 of a brand channel department's "2026 Spring Ordering Handbook" states: "The new-season sample list is subject to Annex 2; Annex 2 contains 12 items, numbered SP-01 through SP-12." The accompanying "Credit Approval Form" states: "Dealer A applies for a credit line of RMB 800,000; approval status is pending regional manager review." A draft sentence intended for internal training reads: "Dealer A's RMB 800,000 credit line has been approved and orders may be placed." A separate store service handover request reads: "Introduce samples SP-01 through SP-12 to Dealer A and confirm order quantities."
Constraints at the citation tracing step
The search node citation output can be empty, and downstream design must accept that possibility rather than treat it as a failure. Because linked bases must share one vector model, the vector model identifier of each candidate base is checked before linking, and bases whose identifiers differ are excluded. This check applies to knowledge bases only and does not extend to business registration sites or other external systems.
The minimum relevance score filters low-relevance results in semantic retrieval or after reranking. It is not a category filter, an entity filter, or a permission filter, and a similarity setting does not exclude unrelated businesses or unauthorized records. A retrieval hit on the credit approval form does not change the form's stated status of pending regional manager review.
Mixed retrieval combines semantic and full-text recall through RRF, and reranking requires a rerank model that is present and enabled. Image retrieval requires an image input and a model with the corresponding capability. These optional paths in the overall retrieval flow are not executed on every request.
Reasoning budget belongs to model thinking configuration. The citation cap is measured in tokens and controls the amount of cited content. The two fields act on different objects and cannot substitute for each other.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge base | Link only the current-season ordering material base with one vector model | Compare vector model identifiers across candidate bases one by one; exclude any base with a different identifier and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Compare numbered-item hit results against semantic retrieval alone and observe whether numbers are recalled by full-text matching |
| Minimum relevance score | Assumed starting point 0.3 | Compare against 0.2 and 0.4 for recalled item count and irrelevant fragment count; record actual observations and treat the value as a starting point, not a default |
| Citation cap | Assumed starting point 3000 tokens | Compare against 2000 and 4000 for citation completeness; measure in tokens and do not convert to an item count |
| Result reranking | Enable when a rerank model is available | Compare candidate ordering with reranking off; if no rerank model is configured, do not enable it |
| Citation display check | Compare each answer sentence against the cited source text | Use page 3 of the sample handbook and the credit approval form text as comparison objects; answer fluency does not replace source evidence |
Three common errors
- Treating a retrieval hit as an approval. The credit approval form states "pending regional manager review." Retrieving that form does not change the pending status. A pending request is not approval, and pending status does not by itself show that an operation is unauthorized.
- Linking knowledge bases with different vector models. Before linking, check each base's vector model identifier and exclude bases whose identifiers differ. Do not run cross-model retrieval to compare missed recalls.
- Treating the reasoning budget as the citation cap. The reasoning budget belongs to model thinking configuration, while the citation cap controls cited content in tokens. The fields and their objects differ, and neither replaces the other.
How to verify the setup
- Input the page 3 text of the sample handbook and the draft sentence "Dealer A's RMB 800,000 credit line has been approved and orders may be placed." Compare the approval status stated in the draft sentence with the credit approval form text "pending regional manager review." Both refer to the same subject and the same period, and the same approval status field carries different values, which forms a contradiction. The draft sentence is not supported and should be rewritten as pending review.
- Input the store service handover request "Introduce samples SP-01 through SP-12 to Dealer A and confirm order quantities." Compare the number range in the request with the handbook Annex 2 text "12 items, numbered SP-01 through SP-12." When the ranges match, the request may proceed to manual review. Whether it is executed still depends on the sample list version and the authorization status.
- Input an empty citation array [] and compare whether the downstream answer step accepts an empty result. This check confirms downstream design and does not claim that an actual retrieval necessarily returns an empty array. If an actual retrieval returns non-empty citations, compare each citation text against its source version one by one.
Where a material field is not provided, the value remains unknown. Where only related-stage evidence exists, the matter remains unsupported or pending verification. Signing, cooperation reaching agreement, and contract taking effect are three distinct facts.