What the documents look like
Franchise filing management works with legal disclosure and the commerce department system. It does not handle individual site engineering. For the inventory procurement and supplier materials task, the relevant materials are the franchise contract, the operations manual, filing materials, and information disclosure documents. These are maintained by the franchisor headquarters, franchise legal staff, and the commerce liaison, and they update irregularly.
Supplier names, purchase batch references, store lists, and contract ledger entries can appear in the same body of material. The vector model and index step has to turn these materials into searchable vectors, and associated knowledge bases inside one search must use the same vector model. Whether material has been trained and whether an index has been generated determine whether a search can recall the corresponding source text.
A further constraint comes from the filing context itself. A search hit returns source content. It does not return an approval, an authorization, or a completed business action. Procurement material can describe who may replace a supplier, while the actual replacement request sits in a separate approval process. The retrieval layer carries no permission or entity isolation, so a recalled passage about supplier replacement rules says nothing about whether a specific replacement request was approved.
Constraints at the vector indexes step
In FastGPT, a knowledge base is made of libraries, collections, and data. The smallest search unit is the library. Vector storage uses the PostgresSQL PG Vector plugin with an HNSW index. MongoDB stores source data in the dataset.datas table, and an indexes field records the corresponding vector IDs, so one data item can map to multiple vectors. Image retrieval depends on whether an image description index or image vector index was generated at ingestion, and on whether the vector model supports image input.
Semantic search relies on vector distance. Full-text search suits numbers, models, and proper nouns. Hybrid search fuses multiple recall paths through RRF, and Rerank performs a second ordering of candidate text. The minimum relevance value takes effect only in semantic search or when result reranking is used. The reference cap is controlled in tokens.
Retrieval does not carry permission or entity isolation. Before associating knowledge bases, the vector model identifier of each candidate library must be checked for consistency. Libraries with different vector models cannot be associated together; the check is to compare identifiers, exclude those that do not match, and record what was observed. A cross-model search is not run to compare missed recalls.
The overall retrieval flow lists optional paths, and not every request executes all of them. Image retrieval requires a corresponding image input and model capability. Reranking requires a rerank model and must be enabled according to settings. The flow diagram does not imply that every request passes through every stage.
Configuration and implementation checks
The table lists only settings or implementation checks directly relevant to this page's task.
| Configuration or check | Value or action in this page's sample | Basis and explanation |
|---|---|---|
| Associated knowledge base vector model | First compare the vector model identifiers of the libraries to be associated | Associated knowledge bases must use the same vector model |
| Index content | Segment the framework note into supplier directory, qualification archiving, and replacement restriction sections for indexing | Complete and focused index content supports precision |
| Multi-vector mapping | Add one additional index group for the paragraph containing the replacement restriction | One data item can map to multiple vectors; a hit on any one recalls the item |
| Image index | The sample has no image input, so no image index is generated | Image retrieval depends on whether a corresponding index was built at ingestion |
| Search mode | Number-type supplier queries can use full-text or hybrid search | Full-text search suits numbers, models, and proper nouns |
| Reference cap and minimum relevance | This hypothetical check starts at a reference cap of 2000 tokens and a minimum relevance of 0.4 | This starting point is chosen for this paper check and must be compared against actual recalled source text from the same question set before adjustment |
Three common errors
- Treating a search result as an authorization basis. In the sample, the handover request has not received any approval record. Recalling the sentence about franchisees not replacing suppliers on their own shows only what the material says; it does not show that the request was approved. To check, input the request text, compare the search output against the source data block, and observe whether any approval record appears. If none appears, record it as pending verification.
- Associating knowledge bases across different vector models. If two libraries have different vector model identifiers, they cannot be associated. To check, input the two model identifiers, compare whether they match, and observe that a mismatched library is excluded. A cross-model search is not run to compare missed recalls.
- Treating RRF as a switch that can be turned on. RRF is the fusion process after hybrid search, not an independent configuration item. To check, input the same question through semantic search and full-text search separately, compare the two candidate sets, and observe whether the fused references draw from both paths. This observation does not support a claim that any particular item will rank first.
How to verify the setup
- Verification one: input the sample source sentence "This note applies to the food ingredient supplier directory for East China stores in 2025," compare whether the returned reference content contains that sentence and the phrase about franchisees not replacing suppliers on their own, and observe whether that source data block appears within the reference cap. If it does not appear, record it as not recalled and do not infer a cause.
- Verification two: input the text of the handover request to be submitted, compare the search result against the approval record field, and observe whether any approval status appears. If none appears, record it as pending verification and do not treat the search hit as authorization.
- Verification three: input the vector model identifiers of the two knowledge bases to be associated, compare whether the identifiers match, and observe that association may continue only when they match. When they do not match, exclude the library and record the actual observation. For empty-result downstream handling, a paper input of [] can be used to check the downstream design; this page does not claim that a rule search returns [].
The hypothetical drill sample used across this page: a franchisor headquarters legal staff member prepares an East China supplier procurement framework note in March 2025. The source text reads: "This note applies to the food ingredient supplier directory for East China stores in 2025. Supplier qualification documents are archived by headquarters procurement. Franchisees may not replace suppliers on their own." A separate handover request in the same batch reads: "Please add the above note to the filing knowledge base so that franchise legal staff can search for the basis of supplier replacement during contract disputes." This request has not received any approval record and does not attach the original supplier qualification documents.