What the documents look like
Category access review happens before listing or store onboarding, and the review object is the operating entity together with its authorization chain. The material set for inventory procurement and supplier files consists of business licenses, trademark registration certificates, brand authorization chains and industry permits. This page covers one step only: parsing supplier-provided material into retrievable text chunks so that onboarding specialists can check the entity, the object, the validity period and the scope of authorization.
The sources named in the brief are the platform recruitment pages, the national enterprise credit information publicity system, trademark office data and permit systems. Material is collected again for each onboarding, and update timing is irregular. The difficult part of this step is that an authorization chain usually consists of several documents, and a single authorization letter binds a brand, an authorized party, an authorized product scope and a validity period at the same time. If parsing and chunking separate a conditional clause from the party it constrains, retrieval can return an authorization scope while losing the authorized party, and a reviewer working from that result can reach a wrong conclusion. Retrieval returns reference content; it does not establish that authorization exists.
Constraints at the document chunking step
In FastGPT the knowledge base consists of libraries, collections and data. The smallest search unit is the library, and collections serve only to classify data. Vectors are stored in a PostgresSQL vector field with an HNSW index, while MongoDB stores the original data in the dataset.datas table, where an indexes field records the corresponding vector IDs, so one data group can correspond to multiple vectors. At retrieval time the system recalls vectors first, then uses the vector IDs to fetch the original data from MongoDB; results that map to the same data group are merged and the highest vector score is kept.
This structure means chunking decisions directly affect review outcomes. A long authorization-chain text can carry several vectors, and recall of any one of them brings back that data group. Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, hybrid retrieval fuses both through RRF, and ReRank produces a 0-1 score. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filtering threshold and does not perform entity isolation or permission checks.
Hypothetical paper test case. The material text to be checked reads: "Party A, XX Brand Management Co., Ltd., authorizes Party B, XX Trading Co., Ltd., to sell registered trademark goods of class XX on a certain e-commerce platform during the period from 1 January 2025 to 31 December 2025. This authorization does not include sub-authorization." The draft sentence to be assessed reads: "Party B has obtained the brand's sales authorization on the platform, and the authorization scope includes sub-authorization." In this sample the entity, the object, the period and the exclusion are all present in the source text, and the draft sentence takes a mutually exclusive value on the same field, so it is unsupported.
Configuration and implementation checks
| Configuration or check item | Setting relevant to this page | Explanation and boundary |
|---|---|---|
| Search mode | Semantic retrieval, full-text retrieval or hybrid retrieval | Authorization chains contain entity names and numbers, so semantic and full-text observation can be used together |
| Reference limit | Maximum tokens referenced per search | Hypothetical trial starting value 2000 tokens, chosen because the sample authorization letter is short, compared with 1000 and 4000 to observe recall completeness |
| Minimum relevance | A 0-1 value, effective only under semantic retrieval or result reranking | Hypothetical trial starting value 0.3, compared with 0.2 and 0.5 to observe the proportion of irrelevant fragments |
| Question optimization | Coreference resolution and question expansion before retrieval | Can be enabled for follow-up questions about an authorization period, and left off for a single clear question |
| Result reranking | A ReRank model reorders candidates | Requires an available rerank model enabled in settings; without it no rerank score is produced |
| Chunking check | Inspect paragraph completeness against document structure | Manual check that conditions, periods and the constrained party stay in one chunk |
Three common errors
- Treating a retrieval hit as established authorization. The input is the sample source text and the draft sentence; the comparison is the value on the same field; the observed result is that the source text states the authorization does not include sub-authorization while the draft sentence states it includes sub-authorization, so the two are mutually exclusive and the draft sentence is unsupported. Retrieval returns reference content only, and whether authorization exists requires separate verification.
- Treating a successful upload as an updated index. The input is a newly submitted permit file from a supplier; the comparison is training and index status; the observed result is that the upload completes while retrieval still returns the older version, so the item is recorded as pending verification rather than as a newly effective permit.
- Using a similarity setting to exclude unrelated entities. The input is an authorization letter from another company; the comparison is the retrieved result against the target entity name; the observed result is that a low-relevance filter only reduces low-scoring fragments and does not exclude records by entity or permission, so exclusion by entity name remains a manual step.
How to verify the setup
- For the hypothetical paper test case, input the draft sentence "Party B has obtained the brand's sales authorization on the platform, and the authorization scope includes sub-authorization", compare the source text "This authorization does not include sub-authorization" with the draft sentence on the same field, and observe that the two values are mutually exclusive. The result is recorded as unsupported, with the evidence gap retained.
- For the sample source text, input "31 December 2025" together with "authorization period", and compare whether the recalled fragment contains both the period and the constrained party. The condition is met when the entity, the object and the period appear within the same data chunk, which is recorded as complete chunking; if the period and the party fall in different chunks, the item is recorded as pending adjustment.
- For the empty-reference case, input a clearly stated paper retrieval result of [] and compare the downstream design for reference trimming and answer generation. The condition is met when the downstream treats it as empty reference and indicates that no usable content is available; no claim is made that a fabricated rule search necessarily returns an empty array.
Libraries built on different vector models cannot be associated together. Check the library identifier first, exclude those that do not match, and record the actual observation rather than running a cross-model retrieval to compare missed recall. Image retrieval requires corresponding image input and model capability, and reranking requires an available model enabled in settings; both are optional paths and do not run on every request.