What the documents look like
Fresh-food concession counters in hypermarkets are frequently operated by on-site suppliers under a sales-share arrangement. Procurement and supplier records therefore span concession contracts, supplier delivery notes, food inbound inspection records, daily sales reports and shrinkage sheets. These materials are produced daily or per delivery batch, and they are used to reconcile inbound documentation against shrinkage allocation. The vector model and index step operates on this document set. Retrieval output is reference material for internal reconciliation; retrieval by itself does not confirm supplier qualification, settlement amounts or return-liability conclusions.
A second characteristic matters here. The same counter can receive text documents, scanned delivery notes and product photographs in one workflow. Whether an image can be retrieved depends on what was indexed at ingestion time, so the document mix directly shapes which retrieval paths are even available.
Constraints at the vector indexes step
In the documented FastGPT knowledge-base structure, a knowledge base is composed of libraries, collections and data. The smallest search unit is the library. Vector storage uses the PostgresSQL PG Vector extension with an HNSW index. The MongoDB dataset.datas table stores the source data of vectors, and an indexes field records the corresponding vector IDs as an array, so one data group can map to multiple vectors. During retrieval, vectors are recalled first, then the vector IDs are used to look up source data in MongoDB. When several vectors resolve to the same source data group, the results are merged and the highest vector score is kept. Knowledge bases linked for retrieval must use the same vector model; libraries built on different vector models cannot be linked together, so the actual model identifier must be checked first and non-matching libraries excluded. Image retrieval additionally depends on whether an image description index or image vector index was generated at ingestion; supplying an image only on the query side is not sufficient.
Hypothetical paper exercise sample. A store fresh-food department retains a concession supplier delivery note whose text reads: "Supplier A delivered 80 kg of chilled pork to this store on 10 March 2025, accompanied by food inbound inspection record number JH-0310, with complete inspector signature." A shrinkage sheet in the same batch reads: "3 kg of chilled pork written off on 10 March 2025, reason recorded as thawing blood-water seepage." The draft sentence to be assessed is: "This supplier's batch has completed all inspection and the shrinkage is borne by the supplier." That draft exceeds what the source text states. The source records only delivery quantity, inspection record number and signature; the shrinkage sheet records only the write-off quantity and reason, and states no liability allocation.
Configuration and implementation checks
| Configuration item | Relevant value or check on this page | Notes and basis |
|---|---|---|
| Vector model | All collections in one library use the same vector model | Linked retrieval requires the same vector model; check the actual model identifier first |
| Index field | indexes records vector IDs and may be an array | One source data group can map to multiple vectors; check the index and the source text separately |
| Index content | Build indexes from complete, focused text | Index length and retrieval coverage must be weighed against the sample |
| Image index | At ingestion, confirm whether an image description or image vector index was generated | Image retrieval depends on the ingestion index, not only the query side |
| Reference cap | For this hypothetical exercise, start at a small number of tokens | Chosen to compare quoted source text under different settings; compare against adjacent values |
| Minimum relevance | A 0–1 threshold effective in semantic retrieval or with result reranking | This value is a filter threshold, not a final ranking rule |
Three common errors
- Treating the draft sentence "the shrinkage is borne by the supplier" as a fact already stated in the source. The source records only the write-off quantity and reason; liability allocation is not stated. The comparison must follow the source wording, and unstated content is recorded as pending verification.
- Linking libraries built on different vector models directly. The model identifier should be checked first, non-matching libraries excluded, and the actual observation recorded. No conclusion about missed recalls or redundancy should be drawn from this.
- Treating a retrieval hit as authorization or a settlement basis. Retrieval returns reference content; supplier qualification confirmation, shrinkage allocation approval and settlement require separate downstream implementation, and retrieval does not carry that role.
How to verify the setup
- Input the delivery note and shrinkage sheet text above, then compare the draft sentence against the fields stated in the source. If the draft contains a liability-allocation statement absent from the source, the sentence is unsupported. If it stays within delivery quantity, inspection record number and write-off quantity, it matches the source.
- Input two knowledge-base identifiers built on different vector models and compare the model fields. If the identifiers differ, exclude that combination and record it. If they match, proceed to check the retrieval configuration.
- Input a pure image query and compare whether an image description or image vector index was generated at ingestion. If no such index was built at ingestion, image retrieval is unavailable; if one was built, then check the query-side model capability. For empty-result handling, a clearly stated paper input of an empty array can be used to check downstream design, without asserting that retrieval will necessarily return empty.
These checks test the named primary task: whether the vector model and index configuration supports procurement and supplier document retrieval for this counter category. They do not test supplier qualification, settlement approval or any business action.