What the documents look like
Brand outlet operations clear seasonal inventory, and pricing sits apart from full-price brand presentation. The working document set for procurement and supplier records includes discount approvals, slow-moving inventory lists, price tags and gross-margin worksheets. Data comes from ERP inventory age, POS sales, outlet mall campaigns and historical discount records, with weekly or monthly updates. The tracing task here is narrower than general knowledge-base search: a handover note or a draft sentence about a purchase batch, a supplier quotation, a discount authorization or inventory-age status must be traceable to a specific material and a specific sentence inside it, with that material's current state stated.
FastGPT stores a knowledge base as library, collection and data. The smallest search unit is the library, so a search runs against a whole library and collections only classify data. A dataset search node can attach one or more knowledge bases that share the same vector model. Its output is an array of citations whose length may be zero, and a zero-length array still means the output path ran through. That distinction matters here: an empty citation array is not a workflow error, and it is not evidence that a supplier approved anything.
Retrieval itself is not authorization. Forms collect values; business verification, approval, database writes and external synchronization each need their own downstream implementation. The minimum relevance value is a filter threshold and does not isolate businesses, entities or permissions. Hybrid search merges multiple recall paths through RRF, reranking needs a rerank model enabled under the settings, and image retrieval needs image input plus the relevant model capability. Knowledge bases with different vector models cannot be attached together, so identifiers are checked first and non-matching libraries are excluded. The citation cap is measured in tokens, and a thinking budget is a separate setting from that cap.
Constraints at the citation tracing step
The dataset search node attaches one or more knowledge bases that share the same vector model. Output is a citation array, possibly empty. Results are handed to an answering step, and the answer is then compared line by line against the cited source text. A retrieved discount approval sheet does not mean the approval passed. A quotation validity date does not mean a supply period. A pending request is not an approval, and an offered option is not an execution.
Hypothetical paper exercise sample. A brand outlet clearance group is organizing procurement and discount materials for a batch of off-season coats in March 2024. Material A reads: "Supplier A quotation dated February 2024 states: style W-221 coat, tax-inclusive unit price 180 yuan, quotation valid until 31 March 2024." Material B reads: "Discount approval form dated 5 March 2024 records: style W-221, requested discount rate fifty percent, approval status pending review." The draft sentence under judgment is: "Supplier A has agreed to supply W-221 at fifty percent, and the discount is in effect." This draft touches subject, object, status and validity scope, and each must be compared against Materials A and B.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Attached knowledge bases | Only procurement and discount libraries sharing one vector model | Compare identifiers against another vector-model library; exclude when identifiers differ |
| Search mode | Hybrid search | Compare with semantic-only search and check whether coded fields are recalled |
| Citation cap | A hypothetical number of tokens | Compare with the citation length the current answer needs and observe truncation |
| Minimum relevance | A hypothetical mid-range value | Compare with recall relevance scores and observe how many entries remain after filtering |
| Result reranking | Enabled when a rerank model is available | Compare ordering with reranking off and observe candidate order changes |
| Citation output handoff | Pass the citation array to the answering step | Compare the answer against each cited passage and confirm the source text is traceable |
Three common errors
- Treating pending review as approved. Material B records approval status pending review, while the draft says the discount is in effect. Check: input the Material B text, compare the field values "pending review" and "in effect", and observe whether the draft matches the source. The draft does not obtain support.
- Treating a quotation validity date as a supply period. Material A states validity until 31 March 2024, which constrains the quotation only and states no supply period. Check: input the Material A text, compare whether "quotation validity" and "supply period" are the same field, and observe whether the draft adds content the source does not state.
- Treating a retrieval result as an authorization record. Retrieving a discount approval form does not mean the approval passed. Check: input the retrieved citation array, compare the array contents with the approval status field, and observe whether recall is equated with approval. The approval status remains pending review.
How to verify the setup
- For the draft sentence "Supplier A has agreed to supply W-221 at fifty percent, and the discount is in effect", input the Material A and Material B texts and compare four fields: subject, object, approval status and validity scope. The sentence obtains support only when Material B shows approval status approved and Material A states a supply period. With the current materials showing pending review, the sentence does not obtain support.
- Input an explicit paper citation array
[]and check whether the downstream answering step is designed to accept an empty array. Observe whether the downstream path still runs. Do not claim that any real search will necessarily return an empty array. - Check the vector-model identifiers of the attached knowledge bases, exclude libraries whose identifiers differ, and record the actual observation. Do not run cross-model retrieval to compare missed recall, and do not infer runtime results from a shared attachment.
This page covers one scope only: brand outlet operations, procurement and supplier records, citation tracing. At an operations handover, the state of each material, action and request must be stated, and retrieval or a form must not be treated as an authorization execution.