What the documents look like
For platform promotion enrollment and price management, the working materials are the activity enrollment form, the promotional price approval form, the inventory lock table, and the post-campaign review table. The center of gravity sits on price, stock, and campaign timing, which separates this work from product title compliance review. Data arrives from platform marketing consoles, activity back ends, ERP inventory, and historical transaction prices, and it refreshes weekly or once per campaign.
Inside product data and selection question answering, these documents become knowledge base content for retrieval and citation. An enrollment form records the enrolling entity, the campaign window, and the product scope. A price approval form records the approval object, the price figure, and the approval status. An inventory lock table records the locked object, quantity, and lock period. A review table records actual figures after the campaign closes. How completely these fields are filled directly affects whether a vector index can match the corresponding content.
Hypothetical paper exercise sample
The following material is used for a paper-based check and points to no real party.
Enrollment form text: "Enrolling entity: a brand sales planning group in the East China region; campaign object: Product A vacuum flask; campaign window: June 1 to June 18, 2025; enrollment status is submitted, awaiting platform review."
Promotional price approval form text: "Approval object: Product A vacuum flask; proposed promotional price: 79 yuan; approval status: department head has signed, finance business partner has not signed; price validity: June 1 to June 18, 2025."
Draft sentence to be judged: "The promotional price approval for Product A vacuum flask is complete, so enrollment for this campaign can proceed at 79 yuan."
The draft sentence treats "department head has signed" as equal to "approval complete," while the approval form states the finance business partner has not signed and the approval status is not complete. Retrieval can return both documents, yet whether approval is finished still depends on the approval status field. A retrieval result does not constitute approval authorization.
Constraints at the vector indexes step
FastGPT builds knowledge bases with an Embedding approach. A smaller vector distance means greater similarity, and retrieval uses topk recall. A knowledge base consists of library, collection, and data. The smallest search unit is the library; collections only classify and manage data. Vector storage uses the PostgresSQL PG Vector plugin with an HNSW index. The MongoDB dataset.datas table stores vector source data, and its indexes field records the corresponding vector IDs, so one data group can map to multiple vectors.
Multi-vector mapping balances content length against semantic richness. A longer text can carry multiple vectors, and recall of any one vector brings back that data. Image retrieval depends on whether an image description index or image vector index was generated at ingestion and on whether the vector model accepts image input. The knowledge base search node requires linked knowledge bases to share the same vector model. That is a linking condition; retrieval itself performs no permission or entity isolation.
The output citation returns as an array and its length can be 0, so the output path still completes when nothing is found.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the promotion material library that uses one vector model | The search node requires the same vector model; check each model identifier and exclude those that differ |
| Search mode | Hybrid retrieval | Enrollment forms carry campaign numbers and entity names, so semantic recall and full-text recall run together and merge through RRF |
| Citation limit | A hypothetical starting point of 3000 tokens for this exercise | This is a paper trial value; compare it against the actual length distribution of candidate chunks before adjusting |
| Minimum relevance | A hypothetical starting point of 0.4 for this exercise | This is a paper trial value; it applies only under semantic retrieval or result reranking, and cited text is reviewed against the source |
| Result reranking | Enable ReRank according to settings | Reranking needs a model and must be enabled by settings; the rerank score is 0-1 and can drive filtering |
| Question optimization | Enable for multi-turn follow-up questions | Question optimization adds one model call before retrieval for coreference resolution and question expansion |
Three common errors
- Treating "department head has signed" as completed approval. The approval form states the finance business partner has not signed, so two field values differ and the same field is incomplete. The draft sentence does not match the source. Compare the approval conclusion in the draft sentence against the approval status field on the form.
- Treating retrieval recall as authorization. Returning the enrollment form only shows the content is relevant to the question. Whether enrollment passes and whether the price can be executed require platform review and the approval process to finish. Compare the business action claimed in the draft sentence against the authorization status recorded in the material.
- Linking knowledge bases that use different vector models. The node requires linked knowledge bases to share one vector model. Exclude a library whose model identifier differs, and do not run cross-model retrieval to compare recall differences. Compare the model identifiers of each library for consistency.
How to verify the setup
- Input the enrollment form text above together with the draft sentence, and compare "awaiting platform review" in the draft against the enrollment status field in the source. If the retrieved citation contains the enrollment form text and the draft sentence does not describe a pending review as passed, the check is satisfied. If the draft describes a pending review as passed, it is not satisfied.
- Input the promotional price approval form text together with the draft sentence, and compare "approval complete" in the draft against the approval status field on the form. The form states the finance business partner has not signed. If the citation preserves that unsigned status and the draft sentence does not conclude that approval is complete, the check is satisfied. If the draft concludes approval is complete on that basis, it is not satisfied.
- Input a retrieval request containing only the enrollment form text, and compare whether the returned citation's entity, object, and window match the source. If the returned citation contains the enrolling entity, campaign object, and campaign window, and no approval conclusion absent from the source appears, the check is satisfied. If retrieval returns no match, record it as unsupported, preserve the evidence gap, and do not infer a system fault from it.
An empty result can be checked against downstream design with an explicit paper input of []. This page does not claim that a retrieval under invented rules must return an empty array.