This documentation covers functionality available in FastGPT v4.17.0.
What this category of data looks like
Data sources for this category include enterprise procurement platforms, public resource trading platforms, supplier databases, contract management systems, and audit issue databases. The update frequency is daily. Documents processed daily cover procurement plans, requirement specifications, bidding documents, bid evaluation records, contracts, acceptance slips, and supplier complaint materials. Each document type includes fields such as procurement method, budget basis, scoring rules, supplier qualification requirements, and performance clauses. There is no unified fixed format, and some documents include attached materials.
What constraints these characteristics impose on knowledge base retrieval and recall workflows
Multi-source data must be divided into datasets by business module to avoid irrelevant cross-category data interfering with review processes. Daily updates require configured automatic synchronization tasks to ensure the timeliness of retrieved data. Diverse document types with significant field differences require tailored segmentation rules for different document types. Core time-consuming steps focus on procurement method verification, scoring rule review, contract performance comparison, and similar tasks. Retrieval must prioritize recalling precise content for target fields. Strict compliance requirements mean retrieved results must strictly match enterprise systems and disciplinary audit standards, to avoid introducing non-compliant content.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Tendering and procurement documents are highly specialized. This range filters low-similarity irrelevant content to ensure retrieval accuracy |
limit | Top 10 results | A single retrieval returns an appropriate number of results, avoiding context overload while covering multiple materials required for core review tasks |
usingReRank | Enabled, recall 20 entries before reranking | Balances retrieval speed and accuracy. First recall basic results, then use reranking to filter highly relevant content |
datasetSearchUsingExtensionQuery | Enabled | For specialized tendering and procurement terminology, optimize queries to expand retrieval scope and match specialized document fields |
collectionIds | Divided by business modules such as procurement center, bid evaluation records, and contract management | Isolate datasets by business scenario to avoid irrelevant data interfering with review workflows |
Three common configuration errors
- Retrieval results return local image file addresses. This occurs when no external link mapping rule for knowledge base attachments is configured. Only local paths are stored, making images unable to display properly in external scenarios.
- Retrieval latency becomes too high after reranking. This occurs when the number of recalled entries is not set appropriately. Too many recalled entries increase reranking calculation load, or a reranking model with high computational overhead is selected.
- System logs show the actual called model does not match the configured model. This occurs when the model for the knowledge base parameter optimization step is not unified, leading to deviations in model calls across the workflow chain.
How to confirm configurations are correctly set
- Check data synchronization logs to confirm that daily updated procurement, bid evaluation, contract, and other documents have been successfully imported into the corresponding datasets.
- Initiate test retrievals for procurement methods and scoring rules, and verify that returned results only contain content from documents within the business scope.
- Check system call logs to confirm that models used for retrieval, reranking, and query optimization match the configured settings.
- Construct review questions not covered by the knowledge base, and verify that the system only returns a no matching results prompt without generating irrelevant responses.
The values provided here are common starting points. They 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.