What this category of data looks like
Data for this category comes from the centralized treasury payment system, unified budget management system, electronic invoice platform, and contract management system. It is updated daily. Documents processed daily include payment applications, fund plans, invoices, contracts, acceptance records, reimbursement approval forms, and payment vouchers. Each document typically includes fields such as budget indicator code, economic classification code, amount, and approval node information. Units involved include RMB yuan, batch number, contract number, and others. Most document structures combine structured forms and unstructured attachments.
Constraints on knowledge base retrieval and recall
Dispersed data sources and daily updates require the knowledge base to support incremental synchronization of multi-system data. This avoids delays caused by full index reconstruction. Daily documents include both structured fields and unstructured attachments. Retrieval must cover both structured metadata and text content. A single audit links to multiple documents. Retrieval must accurately recall contracts, invoices, and acceptance records that match the current declaration form. This prevents irrelevant results from interfering with audits. Since errors directly affect fund compliance, retrieval result relevance must meet audit requirements. This reduces compliance risks from false recalls.
How to configure this
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches compliance requirements for centralized treasury payment audits, balances recall accuracy and coverage, adjust based on actual testing |
limit | Top 8–12 results | Adapts to the number of documents linked to a single audit, avoids context overflow that impairs model understanding |
searchMode | vector | Prioritizes semantically related documents, supplemented by structured metadata retrieval to meet recall needs for multi-source data |
usingReRank | Enabled | Re-ranks recall results, raises the display priority of high-compliance documents to meet accuracy requirements for audit scenarios |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval keywords, covers synonymous expressions for professional terms such as economic classification and budget indicators, adapts to specialized vocabulary in fiscal business |
collectionIds | Specify the dedicated collection for treasury payment audits | Isolates knowledge base data from other fiscal business categories, prevents irrelevant documents from interfering with retrieval results |
Three common configuration errors
Mistake 1
Phenomenon: Significant differences between results from online chat and API calls. Deviations are large when the API enables stream: false and detail: true. Cause: Global knowledge base retrieval parameters are not unified. Parameter override rules vary across call chains.
Mistake 2
Phenomenon: API calls return the error invalid configuration parameter name "hnsw.max_scan_tuples". Cause: An invalid configuration parameter name was used, causing the retrieval engine to fail to recognize the request.
Mistake 3
Phenomenon: After upgrading FastGPT to version 4.9–4.10, the previously indexed knowledge base cannot trigger vector retrieval, and searches return no results. Cause: The old index format is incompatible with the new version’s vector retrieval engine. Reindex or update the index configuration to resolve this.
How to confirm the configuration is complete
- Access the FastGPT knowledge base configuration page, verify that the values of retrieval parameters such as
similarityandlimitmatch the preset configuration. - Initiate a retrieval request matching daily audit scenarios, confirm that the relevance of returned results meets business audit requirements.
- Initiate the same retrieval request via online chat and API call separately, confirm that both return consistent retrieval results.
- Import a test structured document, confirm that retrieval results properly associate corresponding fields and content.
The values provided here are common starting points and should be measured 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.