What the documents look like
Category access happens before listing or onboarding. The core materials are the applicant's legal identity and the chain of authorization. A business license shows the applicant entity and its registered business scope. A trademark registration certificate shows who holds the right. A brand authorization chain shows the path from the rights holder to the applicant. An industry license shows the conditions for a specific category. These materials are updated at irregular intervals, and each onboarding submission is a separate cycle. Document status moves among submitted, supplement requested, and rejected.
For product questions and selection questions, the retrieval target is the chunked content and indexes built from those materials. In FastGPT, a knowledge base contains multiple collections, and a collection contains multiple groups of data. The smallest search unit is the knowledge base. One group of data can map to multiple vectors, and the indexes field records an array of vector IDs. The original text of a material and the index content are checked separately. Retrieval precision improves when index content is complete and focused on a single meaning, and recall coverage improves when the number of indexes increases.
Constraints at the vector indexes step
Knowledge bases that are associated for retrieval must use the same vector model. Before association, check the actual model and its supported input types. Image retrieval also depends on whether an image description or image vector index was generated at ingestion time; uploading an image on the query side is not sufficient. The same-vector-model requirement applies only to associated knowledge bases. External data sources such as the national enterprise credit information system and license systems each have their own access method, and this requirement does not extend to those external systems.
Semantic retrieval relies on vector distance, full-text retrieval relies on keyword matches, hybrid retrieval merges both paths through RRF, and reranking performs a second sort on candidate text. The minimum relevance value is a filter threshold. It takes effect only in semantic retrieval or when result reranking is used, and it does not perform permission or entity isolation. When retrieval returns nothing, the reference output is still an array and its length can be 0.
Hypothetical paper exercise sample: the original text of a submitted material reads, "The business license states the entity is Jia Trading Co., Ltd., and the registered business scope includes prepackaged food sales. The trademark registration certificate states the Class 30 trademark holder is Yi Food Factory. The authorization chain document states that Yi Food Factory authorizes Jia Trading Co., Ltd. to use the trademark from January 1, 2025 to December 31, 2025, with an authorization scope of online sales." A draft sentence to be judged reads, "Jia Trading Co., Ltd. has obtained authorization from Yi Food Factory for the Class 30 trademark for all of 2025 for online sales." The handover request reads, "Please retrieve and confirm whether this entity holds the qualification for this category."
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only the qualification materials base is associated | Its vector model identifier must match that of the selection question base; check identifiers before associating |
| Search mode | Hybrid retrieval | Certificate number questions use full-text retrieval, natural language questions use semantic retrieval, and the sample contains both |
| Minimum relevance | Starting point 0.5 | This is a hypothetical starting value in the 0-1 range; compare referenced original text against the same question set at 0.3 and 0.7 |
| Reference limit | Starting point 2000 tokens | Reference volume is controlled by tokens; compare reference completeness against 1000 and 4000 |
| Question optimization | Enabled | Enabled for multi-turn follow-up scenarios; compare retrieval results with the disabled state |
| Result reranking | Enabled | Second sort when candidates are numerous; compare ordering results with the disabled state |
Three common errors
- Treating a retrieval hit as a verified qualification. Retrieval returns chunks of material content. Whether the entity qualification is valid and whether the authorization chain is complete require manual comparison against the original text. With the sample material as input, compare the retrieved reference against the draft sentence. A retrieval hit does not change the authorization state.
- Associating knowledge bases that use different vector models. The input is the qualification materials base and the selection question base. Compare their vector model identifiers. When the identifiers differ, exclude that base. Do not run cross-model retrieval to compare missed recall.
- Treating a pending review status as approved. The input is the handover request and the material status field. Compare the status values. Pending review, submitted, and approved are distinct values. A pending review does not demonstrate that operating permission is absent.
How to verify the setup
- Input the sample material original text and the draft sentence. Compare the entity, object, period, and scope in the draft sentence against each field in the original text. The original text states the period as January 1 to December 31, 2025, and the scope as online sales. The draft sentence states all of 2025 and online sales authorization. The period wording and the scope wording require field-by-field checking. If the draft sentence contains an offline sales authorization that the original text does not state, that field is unsupported.
- Input the handover request. Compare the retrieved reference content against the entity named in the request. The retrieved reference must fall within the target retrieval scope to which access is held. When the reference is empty, carry it downstream as an array of length 0 and record the gap. Do not infer a qualification conclusion.
- Input the same set of qualification questions. Compare the referenced original text before and after adjusting the vector model or index. Observe whether the referenced content changes. Do not judge the effect from the vector score alone, and do not guarantee a first-position hit or full recall.