What the documents look like
Platform listing compliance covers the truthfulness of promotional text and the display of qualification materials. It does not cover parcel delivery. For inventory procurement and supplier records, the material set includes supplier qualification files, purchase contracts, quality inspection reports, prohibited-term lists for advertising, and product title tables. These materials enter the knowledge base when a new item is listed or when a record changes, and they are later retrieved for cross-checking.
Two properties of this material shape the chunking step. First, each parsed result has to point back to the document identity and version it came from, so a reviewer can trace a retrieved fragment to a specific file. Second, a chunk has to keep the relationship between a condition and the content that condition governs. A licence validity date, a certificate number and the issuing date belong with the supplier they describe.
A hypothetical drill sample is used throughout this page. Supplier qualification file, original text: "Supplier A, business licence valid until 30 June 2027, food business licence number XK-2024-0871, quality inspection report for this batch number QC-2024-1120, issue date 5 November 2024." The sentence under review is: "Supplier A holds complete qualifications, valid long term." This sentence states no specific validity date and no licence status. It does not match the original wording "valid until 30 June 2027" and remains unsupported by the source text.
Constraints at the document chunking step
Document parsing and chunking turns raw material into retrievable data. In the FastGPT structure of library, collection and data, the smallest search unit is the library; collections only classify and manage data and do not affect search results. Original data is stored in the dataset.datas table in MongoDB, where the indexes field records the corresponding vector IDs, and one set of data can map to several vectors. Vectors are stored in a vector field in PostgresSQL with an HNSW index. During retrieval the vectors are recalled first, then the vector IDs are used to fetch the original data from MongoDB; entries that map to the same original data are merged and the highest score is kept.
Segmentation should preserve structural and semantic completeness. A longer text can carry several vector groups, and recall of any one group recalls that data. Retrieval can combine semantic search, full-text search, hybrid search, question optimisation and reranking; hybrid search merges results through RRF. The reference limit is counted in tokens. The minimum relevance value applies only to semantic search or to reranking, and both are filter conditions. Neither carries permission or entity isolation. Libraries built on different vector models cannot be associated together.
For procurement material, this means a certificate number and its validity period should sit in the same chunk so that a precise query for the number reaches the surrounding conditions. A chunk that separates the licence number from the validity date weakens later comparison against the source document.
Configuration and implementation checks
| Configuration item | Setting or check for this page | Implementation note |
|---|---|---|
| Search mode | Semantic search, full-text search or hybrid search | Choose by question type; identifier queries favour full-text or hybrid |
| Reference limit | Counted in tokens | The document gives no specific value; check against actual candidate length |
| Minimum relevance | A 0-1 value, effective only for semantic search or reranking | The document gives no default; adjust against observed results |
| Question optimisation | Coreference removal and question expansion before retrieval | Enable as needed; adds one model call |
| Result reranking | ReRank second-pass ordering, score 0-1 | Requires a rerank model enabled in settings |
| Image retrieval | Vision model or image-capable vector model | Requires image input and model capability; not run on every request |
For the drill sample, a search for "QC-2024-1120" is an identifier query, so full-text or hybrid search is the relevant mode. The reference limit and minimum relevance have no documented numeric value, so any starting figure used in a paper exercise must be labelled as a hypothetical starting value for that exercise, with the reason for choosing it and the value it is compared against. It must not be described as an official default.
Three common errors
- Treating a successful upload as an updated index. After a material change, the actual training and index state has to be checked. The upload action on its own does not show that an index was generated.
- Treating minimum relevance as a permission or entity filter. This value only filters low-relevance results. It does not exclude unrelated businesses or unauthorised records.
- Describing the
RRFmerge in hybrid search as a switch.RRFis the process that merges multiple result paths. It is not a separate configuration item that can be turned on.
How to verify the setup
- Input the drill sample original text and the sentence "Supplier A holds complete qualifications, valid long term." Compare the sentence against the validity field in the original. The original states "valid until 30 June 2027"; the sentence states no such date and says "valid long term". The expected observation is that the sentence is unsupported, and the evidence gap is recorded. This check tests the primary task of comparing proposed copy against source wording.
- Input the identifiers "QC-2024-1120" and "XK-2024-0871" from the sample, once with full-text search and once with semantic search. Compare the two recall paths. Observe whether full-text search reaches the chunk holding the identifier and whether semantic search recalls the same chunk. The result shows how segmentation affects precise queries. A first-hit guarantee is not claimed.
- Input an empty-result handover case using the explicit paper input
[]and check the downstream design. Compare how the downstream handles empty references. Observe whether it treats the input as empty references. No claim is made that a rule-based search must return[]. The missing evidence remains missing.
Across these checks, a pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Retrieval does not grant authorisation, and a relevance score is not a category, entity or permission filter.