What the documents look like
A hypermarket fresh-food concession counter is often run on site by a supplier, with revenue shared according to sales. Product reference and selection questions therefore have to work across four document families at once: the concession contract, the daily sales report, the shrinkage (loss) sheet, and the food purchase inspection record. These materials originate from different parties and carry fields belonging to the hypermarket fresh-food department, the concession supplier, and the store manager in the same file. Daily sales reports and food purchase inspection records are produced daily or per batch, so the material changes frequently. If parsing and chunking separate a conditional clause from the text it constrains, a reader checking the material can misread the scope of application. Food-safety penalties, gross-margin disputes, and customer claims are the downstream consequences of error, so each chunk should retain the document identity and version that lets a reader trace back to the source text.
Constraints at the document chunking step
In FastGPT a knowledge base is composed of a library, collections, and data. The smallest search unit is the library; collections only classify data and do not affect search results. Vector retrieval uses the PostgresSQL PG Vector plugin with an HNSW index. The MongoDB dataset.datas table stores the original data, and its indexes field records the corresponding vector IDs, so one set of data can map to multiple vectors. At retrieval time vectors are recalled first, then the original data is located in MongoDB by vector ID; when records correspond to the same original data they are merged and the vector score keeps the highest value. Chunking should follow the structure of each specific material so that conditions, explanations, and the content they constrain stay together.
Semantic retrieval, full-text retrieval, hybrid retrieval, result reranking, and question optimization are optional paths. Image retrieval requires image input and a model with the corresponding capability; reranking requires a model and must be enabled in settings. RRF is the fusion process for hybrid results. Similarity and relevance do not perform permission or entity isolation, and the requirement that a retrieval library share one vector model does not extend to external systems.
Hypothetical paper exercise sample: the concession contract text supplied for checking reads, "Party A is the fresh-food department of a hypermarket, Party B is a vegetable supplier, and the two parties operate the fresh-food concession counter on the third floor of a store under a sales-share arrangement; the contract period runs from a stated start date to a stated end date, and shrinkage allocation is settled according to the monthly inventory count result." The draft sentence to be judged reads, "The supplier has completed this month's shrinkage allocation settlement." The source text states only the settlement method and the contract period. It does not state whether this month's settlement has been completed, so the draft sentence is not supported.
Configuration and implementation checks
| Configuration or check item | Setting for this page | Explanation and boundary |
|---|---|---|
| Knowledge base and collections | Create collections by material type; search operates at library level | Collections only classify; they do not affect search results |
| Vector retriever | PostgresSQL PG Vector, HNSW index | Used for vector retrieval only |
| Original data storage | MongoDB dataset.datas | The indexes field records vector IDs |
| Multi-vector mapping | One set of data can map to multiple vectors | Merged score keeps the highest value |
| Search mode | Semantic, full-text, or hybrid retrieval | Hybrid retrieval is fused through RRF |
| Reference limit | Measured in tokens, not top k | A hypothetical trial starting value of 2000 tokens is chosen because it makes truncated long chunks observable, and is compared with a 1000-token result |
Three common errors
- Treating "has completed this month's shrinkage allocation settlement" in the draft sentence as a fact already stated in the contract source. The check inputs the contract source text and the draft sentence, compares the four fields of subject, object, period, and status, and judges a contradiction only when the source text states a mutually exclusive value in the same field; when the source does not state it, the result is recorded as unsupported.
- Treating a successful upload as proof that retrieval has been updated. The check inputs the changed material and one terminology query, compares the actual training and index state, and records a pending review when retrieval still returns the old content.
- Using a similarity score to exclude unrelated business or unauthorized records. The check inputs one natural-language query and compares the returned content against the target retrieval scope. Similarity does not perform permission filtering; unauthorized content is handled through access control.
How to verify the setup
- Input the contract source text and the draft sentence from the hypothetical sample, compare the subject, object, period, and status fields, and record the draft sentence as unsupported when the source text has no corresponding statement. Meeting this condition is a pass.
- Input three query types (terminology, identifier, and natural language) and compare whether chunking keeps conditions together with the content they constrain. A condition separated from its text is recorded as a failure branch.
- Input an explicit paper-based empty result
[]and check the downstream design against it, comparing whether the downstream handles it as an empty reference. Do not claim that a fictional-rule retrieval necessarily returns[]. Retrieval observation records only what is actually returned and does not guarantee a first-hit match or full recall.
This page addresses only the hypermarket fresh-food concession counter, product reference and selection questions, and document parsing and chunking. Whether materials, operations, or requests are in a given authorization state during handover is not established by retrieval or by a form.