What the documents look like
An MCN agency manages long-term compliance for multiple creator accounts rather than serving a single brand campaign. The procurement side of that work draws on the MCN back office and the contract ledger, and the materials involved include creator signing contracts, training records, violation handling forms, and account scheduling sheets. Procurement and supplier materials sit alongside those documents, and they commonly place the supplier name, procurement batch, settlement conditions, validity period, and authorization boundary inside one paragraph. When chunking separates a condition from the content it constrains, a retrieved fragment loses the basis for judgment.
Account scheduling and procurement materials update daily and weekly. After an update, the actual training and index state needs to be checked. Retrieval does not perform authorization judgment, and a retrieval hit is not an approval record.
Constraints at the document chunking step
A FastGPT knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections classify data and do not affect search results. In the MongoDB dataset.datas table, vector source data is stored, and an indexes field records the corresponding vector IDs as an array, so one group of data can map to multiple vectors. PostgresSQL stores vectors in a vector field. During retrieval, vectors are recalled first, then the vector IDs are used to find source data in MongoDB; entries corresponding to the same source data are merged, and the highest vector score is kept.
Chunking needs to preserve document identity and version so that a fragment can point back to the source text. Paragraphs are checked against the structure of the specific material, and conditions, explanations, and the content they constrain are kept together. Semantic retrieval, full-text retrieval, and hybrid retrieval are combined through RRF, and ReRank performs a second ordering. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used.
Hypothetical paper exercise sample: the source text reads, "Supplier A submitted a quotation on March 1, 2025. The quotation is valid until April 30, 2025. The settlement condition for procurement batch P-07 is payment thirty days after acceptance. This quotation does not constitute a supply commitment." The draft sentence to be assessed reads, "Supplier A has committed to supply batch P-07 before April 30, 2025." The assessment requires the same subject, the same period, and the same field. The source text states a quotation validity period and a settlement condition; it does not state a supply commitment. The draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Setting or check for this page | Basis and boundary |
|---|---|---|
| Search mode | Use full-text or hybrid retrieval for supplier names and batch numbers | Numbers, models, and proper nouns suit full-text retrieval |
| Minimum relevance | For this hypothetical exercise, start at 0.3 and compare with 0.5 to observe recall changes | Takes effect only in semantic retrieval or result reranking; it is a filter threshold, not a final ordering rule |
| Reference limit | Set by tokens; top k is not used | Documents control the reference limit by tokens |
| Chunking check | Check whether a conditional sentence and the content it constrains are in the same chunk | Chunking completeness requirement |
| Index check | Check the correspondence between the indexes field and source data | One group of data can map to multiple vectors |
| Post-update review | After a material change, check the actual training and index state | A successful upload does not mean retrieval has updated |
Three common errors
- Treating a quotation validity period as a supply deadline. Input: the source text and draft sentence above. Compare the fields "validity period" and "supply commitment." The source text does not contain a supply commitment field, so the draft sentence is unsupported.
- Treating a successful upload as an updated index. Input: one material change record. Compare the upload time with the actual training and index state. When the two do not match, retrieval can still return an old fragment.
- Treating a retrieval result as an authorization basis. Input: one procurement material retrieval hit. Compare the retrieval hit with approval records. When approval records are absent, the hit is not authorized execution.
How to verify the setup
- Input the hypothetical source text and draft sentence above. Compare whether the source text contains a supply commitment field. When the source text lacks that field, the draft sentence is unsupported, and the condition is met.
- Input the supplier name and batch number P-07, using full-text retrieval and semantic retrieval respectively. Compare whether the returned fragments contain the corresponding source text; if they do, the condition is met. For an empty result, use the paper input [] to check the downstream design, and do not claim that a fictional rule must return [].
- Input the same query before and after a material change. Compare whether the returned fragment is still the old version. If an old version appears, the index has not updated, and the gap is recorded for verification.