What the documents look like
Dealer ordering events serve offline dealer reservations and do not directly manage platform storefronts. The materials handled in store service and internal training are the ordering handbook, the ordering contract, the sample list, and the credit approval. These materials update by quarter or by year, so retrieval faces multiple versions of the same document set. The ordering handbook contains new-product policy and ordering rules. The ordering contract contains signing and effective status. The sample list contains model and quantity. The credit approval contains limit and approval status. The parties named in these materials are the brand channel department, regional dealers, and merchandise planning. Field structures differ widely, and identifiers and proper nouns are common. Retrieval must reach these materials through both a semantic path and a keyword path so that store service and internal training staff can check them.
This page addresses one task only: brand dealer ordering event, store service and internal training, knowledge base retrieval and recall. Whether materials, operations, or requests are in an authorized state at handover cannot be settled by retrieval or by a form.
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 using different vector models cannot be associated together. Check the identifier first, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recall.
Full-text retrieval suits precise queries such as identifiers, models, and proper nouns. Semantic retrieval suits natural-language phrasing. Hybrid retrieval merges the two paths through RRF. The citation limit controls, in tokens, the maximum amount of content cited per search. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was changed incorrectly. Retrieval results do not carry permission or entity isolation. A similarity score does not exclude unrelated business records or unauthorized records.
Optional paths shown in the overall retrieval flow are not executed on every request. Image retrieval requires a corresponding image input and model capability. Reranking requires a rerank model and must be enabled by configuration.
Hypothetical paper exercise sample: On 10 March 2025, a brand channel department issues an "East China Spring Ordering Event Sample List." The original text states: "This sample list covers 120 units of Model A samples and 80 units of Model B samples. The quotation is valid until 30 April 2025. The supply period is governed by the ordering contract." On the same day, a "Credit Approval" record states: "Dealer A applies for a credit limit of RMB 500,000, status pending review." The draft sentence to be judged is: "Dealer A has obtained a RMB 500,000 credit limit and may take delivery against it." This draft sentence treats a pending review as an obtained credit limit and treats the quotation validity date as the supply period.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Basis for the check |
|---|---|---|
| Associated knowledge bases | Associate only the ordering handbook, ordering contract, sample list, and credit approval bases that use the same vector model | Node parameter description |
| Search mode | Compare semantic, full-text, and hybrid retrieval separately per test question; do not presume one mode suits all materials | Search parameter description |
| Citation limit | Unit is token; test with actual materials of differing length to see whether content is complete, then adjust the limit | Search filtering description |
| Minimum relevance | Takes effect only under semantic retrieval or when result reranking is used; label relevant and irrelevant samples before adjusting the threshold | Search filtering description |
| Question optimization | Used for coreference resolution and question expansion; compare original and optimized questions and check whether the subject was changed incorrectly | Question optimization description |
| Result reranking | When a rerank model is available and enabled by configuration, candidates are re-ordered and filtered by score | Result reranking description |
Hypothetical starting value for this exercise: the citation limit is set to 2000 tokens. The reason is that this value is sufficient to cover the complete entries of one sample list, and it is compared against the length of one ordering contract summary. This value is a hypothetical starting point, not an official default.
Three common errors
- Treating a pending-review status as an obtained credit limit. Input the credit approval record whose original text states "status pending review," and compare it with the draft sentence "has obtained a RMB 500,000 credit limit." The observable result is that the draft sentence does not match the status field in the original text. A pending review does not demonstrate the absence of an operating permission, and it does not equal an obtained credit limit.
- Treating the quotation validity date as the supply period. Input the original text "the quotation is valid until 30 April 2025; the supply period is governed by the ordering contract," and compare it with the draft sentence "may take delivery against it." The observable result is that the draft sentence merges two distinct facts. A quotation validity date does not indicate a supply period.
- Treating a similarity score as a permission filter. Input a sample record that is semantically close to the question but belongs to another region, and compare the retrieval return with the authorized scope. The observable result is that retrieval may still return that record. A similarity setting does not exclude unrelated business records or unauthorized records, and retrieval does not carry authorization.
How to verify the setup
- Input the credit approval record whose original text states "status pending review" together with the draft sentence "Dealer A has obtained a RMB 500,000 credit limit." Compare the value of the status field. The expected result is that the draft sentence does not match the original status, and the draft must be returned for revision. If the draft sentence retains the wording "pending review," the condition is met.
- Input the original text "the quotation is valid until 30 April 2025; the supply period is governed by the ordering contract" together with the draft sentence "may take delivery against it." Compare the validity date field and the supply period field. The expected result is that the draft sentence lacks a contract basis, and the evidence gap must be marked. If the draft sentence is changed to "the supply period is governed by the ordering contract," the condition is met.
- Input a sample record that is semantically close to the question but belongs to another region. Compare the retrieval return with the manually determined authorized scope. The expected result is that the retrieval result does not constitute authorization and requires separate checking. If the retrieval return matches the authorized scope, record that observation; if it does not match, record the gap and do not claim that retrieval has filtered it. An empty result can be checked against the downstream design using an explicit paper input of []; do not claim that a fabricated rule guarantees retrieval returns [].
This function has been checked against public technical facts for identifying related content in documents; the body text is written independently and is not required to reproduce a fixed five-row table.