What the documents look like
In a multi-store matrix, product material is spread across platform merchant backends, ERP channel pricing, brand anti-counterfeiting systems, and distributor reports. Four document families carry most of the burden for product and selection questions. A store authorization list records which stores hold a sales permit. A price control table records the price boundary for each channel. A diversion complaint form records cross-store circulation disputes. A monthly sales report records actual shipments. These materials update daily and monthly, and the same product can carry different authorization status and price constraints across stores.
For product and selection questions, retrieval has to return the fragments that match the asking subject, the object, and the period. The retrieval result supplies quoted content only. It does not confirm authorization, and it does not execute any approval. A store name and a model number appearing in a returned fragment do not mean the store is permitted to sell that model at that price.
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 must be checked first, non-matching bases excluded, and the actual observation recorded. Retrieval treats the whole knowledge base as the smallest search unit; a collection is only a classification layer and does not affect search results.
A single retrieval can combine question optimization, semantic search, full-text search, hybrid search, image retrieval, and result reranking. These paths are enabled according to the input content and the search parameters. The documented process diagram lists the available paths; it does not mean every request runs all of them. Image retrieval requires an image input and the corresponding model capability. Reranking requires a rerank model and must be enabled by configuration.
The citation limit is measured in tokens, and its length can be 0, meaning the downstream chain still runs even when nothing is found. The minimum relevance value takes effect only under semantic search or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation. RRF is the merge step that combines multiple retrieval paths; it is a process inside retrieval, not a switch. The rerank score falls in the 0-1 range and represents how relevant the retrieved content is to the question.
Assumed drill sample: the source text reads, "The authorization list states: Store A is permitted to sell Model X during the period from January 1, 2025 to December 31, 2025. The price control table states: the minimum retail price for Model X at Store A during the same period is 199 yuan." The draft sentence to be judged reads, "Store A currently holds an indefinite sales authorization for Model X across all channels and may set prices independently." The comparison subjects are Store A, Model X, the 2025 annual period, the authorization period field, and the price field.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select knowledge bases using the same vector model; check the vector model identifier | The authorization list and the price control table must sit in bases with the same vector model |
| Search mode | Choose semantic, full-text, or hybrid search according to the question type | Model numbers and store names suit full-text search; natural-language questions suit semantic search |
| Citation limit | Unit is tokens; test content completeness with materials of different lengths | A hypothetical starting value for this drill covers the authorization list and price table fragments, compared against a shorter value that returns only a single sentence |
| Minimum relevance | Takes effect only under semantic search or result reranking | First label relevant and irrelevant fragments in this sample, then compare threshold effects |
| Question optimization | Used for coreference resolution and question expansion; compare the original question with the optimized question | In a follow-up such as "until when is this store authorized," check whether the subject is altered incorrectly |
| Result reranking | Requires a rerank model enabled by configuration; score range 0-1 | When candidate fragments are numerous, compare ordering before and after reranking |
Three common errors
- Treating a retrieval hit as authorization confirmation. The input is the draft sentence and the authorization list source text. The comparison checks whether the quoted content returned by retrieval contains the authorization period and price fields. The observable result is that only fragment text appears, with no authorization effective status, which shows retrieval has not performed authorization verification.
- Associating knowledge bases built on different vector models. The input is the vector model identifiers of two bases. The comparison checks whether the identifiers match. The observable result is that a mismatch leads to exclusion, and no cross-model retrieval is run to compare missed recall.
- Treating minimum relevance as a permission filter. The input is a question about an unrelated store and already-labelled irrelevant fragments. The comparison checks whether returned content can still include irrelevant fragments after the threshold is adjusted. The observable result is that the relevance score does not recognize store authorization ownership, which shows this setting does not carry entity isolation.
How to verify the setup
- Input the assumed sample draft sentence and the authorization list source text. Compare the "indefinite" claim in the draft sentence with the "January 1, 2025 to December 31, 2025" period in the source. Observe whether the quoted content presents the period field. When the period field is not stated, or only evidence for a related stage exists, record it as pending verification rather than as a negation or a conflict.
- Input the assumed sample draft sentence and the price control table source text. Compare the "may set prices independently" claim in the draft sentence with the "minimum retail price of 199 yuan" in the source. Observe whether the quoted content presents the price field. A contradiction exists only when the same subject, the same period, and the same field carry mutually exclusive values.
- Input a paper-based empty-result scenario, set the quoted content to an empty array, and check whether the downstream design accepts an empty array. This scenario is a paper check and does not claim that a fictional rule retrieval necessarily returns an empty array. Authorization status, approval, and database writes each require their own downstream implementation; retrieval and forms do not constitute authorization execution.