What the data for this category looks like
Data sources for this category include enterprise procurement platforms, public resource trading platforms, supplier databases, contract management systems, and audit issue repositories. Daily processed documents cover procurement plans, requirement specifications, bidding documents, bid evaluation records, contracts, acceptance documents, and supplier complaint materials. Core fields included are procurement methods, budget basis, scoring methods, supplier qualification conditions, and performance clauses. Data updates occur daily. Document length varies widely: from a few dozen characters for acceptance documents to several thousand characters for formal bidding documents.
What constraints these characteristics impose on knowledge base retrieval and recall
Daily updated data sources require retrieval configurations to support incremental synchronization, to avoid reindexing old data. Wide variation in document length requires retrieval chunking to adapt to different text lengths, to avoid merging independent paragraphs or over-splitting long documents. Multiple sources of compliance-related documents require retrieval scope to be limited to the dedicated procurement dataset, to avoid mixing unrelated business documents. Strict compliance review requirements mean retrieval result accuracy and relevance must be prioritized.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the accuracy requirements for core compliance texts such as supplier qualification conditions and scoring methods |
limit | Top 10 entries | Controls the number of recalled results, to avoid context overload impacting model inference |
usingReRank | Enabled | Performs secondary ranking on recalled results for long documents such as bidding documents and bid evaluation records |
datasetSearchUsingExtensionQuery | Enabled | Supplements implicit query terms related to procurement compliance, to optimize retrieval matching accuracy |
collectionIds | Bound to the dedicated procurement dataset | In FastGPT v4.17.0, bind the corresponding collection ID via the dataset management interface to limit retrieval scope |
Three common mistakes
- Phenomenon: When a Chinese query triggers retrieval, results do not include English procurement compliance documents stored in the knowledge base. Cause: English procurement-related documents are not added to the bound dataset collection, or multi-language retrieval configuration is not enabled.
- Phenomenon: After custom delimiters are set, chunking results either merge multiple independent procurement document paragraphs, or split a single long bidding document into overly small chunks. Cause: Chunking parameters are not adjusted based on procurement document structure, or delimiters do not match actual line break rules within documents.
- Phenomenon: The number of retrieved results is far lower than expected, or an error with no results occurs. Cause: The
similaritythreshold is set too high, filtering eligible weakly matched compliance documents, or the corresponding data source collection is not correctly bound.
How to confirm configurations are properly set
- A Chinese query related to procurement compliance supported by FastGPT v4.17.0 may be manually entered, and returned results checked for inclusion of documents such as bidding documents and bid evaluation records from the bound dataset.
- The
similaritythreshold may be adjusted to the low end of the recommended range, and the number of recalled results verified to match the expected scope. - After enabling
usingReRank, recalled result rankings before and after enabling may be compared, to confirm that long texts have more reasonable relevance ranking. - The dataset binding list may be checked to confirm all relevant data sources such as procurement plans and supplier databases are included.
The values provided are common starting points, and should be evaluated against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.