What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, so the daily sales reports, shift schedules, display photographs and stock transfer orders produced in those stores remain with the brand rather than passing through a franchise agreement. Copy drafting for products and marketing draws on this internal material, which means the retrieval scope sits inside that body of records. Daily sales reports and shift schedules are updated by day and by week, while display photographs and transfer orders appear as store activity occurs. Material length varies widely, and the same product can be described differently from one store to the next. The drafting task usually turns on whether a specific promotional sentence has support in the material, so the retrieval step has to surface the report for the right date and the photograph for the right store, leaving the actual comparison to a person.
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. Retrieval only addresses the question of which candidate content to surface. It does not perform permission isolation or entity filtering, and a similarity score does not distinguish store ownership or determine whether a record belongs to the current store. A single retrieval combines semantic search, full-text search, image search, question optimization and reranking according to the input content and search parameters, then produces reference content after similarity filtering and a reference limit trim. The output is an array and its length can be zero, so the downstream path still runs when nothing is found. Image search requires image input together with a vision model or a vector model that supports images, and reranking requires a model and must be enabled in settings. These paths do not execute on every request.
Hypothetical walkthrough sample: the source material reads "From 1 June to 7 June 2024, the daily sales report for Store A records average daily sales of 12 units of a certain summer short-sleeve shirt", "During the same period, a display photograph from Store A shows the short-sleeve shirt in the main display position at the entrance", and "On 3 June, a stock transfer order for Store A shows 30 units of the short-sleeve shirt transferred in". The draft sentence under review is "In its first week on sale at Store A, this short-sleeve shirt became a flagship bestseller". Checking confirms that the material covers only sales volume, display position and transferred quantity for 1 June to 7 June, and states no basis for "first week on sale", "flagship" or "bestseller". The draft sentence is unsupported.
Configuration and implementation checks
The table lists only settings already described in the official documentation or implementation checks relevant to this feature and this drafting task, tied to the sample above.
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and what it is compared against |
|---|---|---|
| Associated knowledge bases | Associate only the product and store material bases that share one vector model | Check the vector model identifier of each base first, exclude any that do not match, and do not perform cross-model association |
| Search mode | Try semantic search and full-text search, then try hybrid search | The sample contains precise terms such as dates and a store identifier as well as a natural-language draft sentence, so the three modes are compared against the test question |
| Reference limit | Set to a token volume that can hold one report plus one photograph description | The unit is tokens; test with material of differing length to see whether content stays complete, then adjust |
| Minimum relevance | Mark relevant and irrelevant samples first, then take a middle value | This value takes effect only in semantic search or when reranking is used, and it is a filtering threshold |
| Question optimization | Enable first for referring expressions such as "this short-sleeve shirt" | Used for coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered |
| Reranking | Enable when candidates are numerous | Reranking produces a 0-1 score that can be used for filtering, and it requires a model |
Three common errors
- Treating retrieval results as authorization results. A draft sentence is entered, a store report is retrieved, and the store data is assumed usable in any copy. Retrieval returns candidate content only; store ownership and scope of use require separate manual checking.
- Treating the similarity score as entity filtering. Raising the minimum relevance is assumed to exclude other stores or unrelated records. That value is a filtering threshold; it does not distinguish stores and does not determine whether a record falls within the current scope.
- Treating an unstated field as a contradiction. The material gives sales only for 1 June to 7 June, and the draft sentence says "first week on sale", which is immediately judged as conflicting with the material. An unstated field, or evidence covering only a related stage, is unsupported or pending verification.
How to verify the setup
- Enter the three source sentences and the draft sentence from the hypothetical sample, compare whether the retrieved reference content covers the three elements of 1 June to 7 June, Store A and the short-sleeve shirt, and observe whether the reference array is non-empty and contains the corresponding source data. If an empty array is returned, record it as not retrieved and check whether the data has completed training.
- Enter the short question "this short-sleeve shirt flagship bestseller" alone, compare the retrieval terms returned before and after question optimization is enabled, and observe whether the subject still points to the short-sleeve shirt and whether the date or store was altered. If the subject is replaced, record that optimization result.
- Enter an attempt to associate a knowledge base built on a different vector model, compare the vector model identifiers of each base, and observe whether any base shows a mismatched identifier. Exclude the mismatched base and record the actual observation without inferring a runtime result from it.