What data for this category looks like
Data for this category comes from the unified budget management system, centralized treasury payment system, accounting system, and asset management system. The update cycle is annual. Daily document output includes department final account statements, balance sheets, income and expenditure final account statements, final account explanations, and government financial reports. The core of these documents is structured tables, paired with written explanations. Fields cover budget codes, income and expenditure details, asset stock, difference explanations, and more. Units use financial measurement standards such as ten thousand yuan, yuan, and others.
What constraints do these characteristics impose on knowledge base retrieval and recall
Annual centralized updates require the knowledge base to support batch structured document imports, to avoid overload during single import operations. The predominantly structured document structure requires the retrieval link to distinguish recall priorities between financial fields and explanatory text, to avoid non-critical information interfering with core queries. Field differences across data from multiple systems require unified mapping before retrieval, to ensure similar data from different sources can be accurately recalled. At the same time, the rigor of final account data requires that recalled results match official statistical standards, to avoid deviations in field units or statistical scope.
How to configure the setup
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the rigor requirements of financial data, filters low-relevance non-final account documents |
limit | Top 8 entries | Controls the number of single retrieval results, avoids context overload affecting financial staff reading |
searchMode | Full-text retrieval | Covers dual retrieval needs for structured fields and unstructured explanatory text |
usingReRank | Enabled | Performs secondary sorting on preliminary retrieval results, improves ranking of highly relevant financial data |
datasetSearchUsingExtensionQuery | Enabled | Automatically supplements synonymous query terms related to final accounts, such as matching "budget execution difference explanation" for "income and expenditure difference" |
collectionIds | Specified by annual final account collections | Limits retrieval scope to only the current year’s department final account document collections, avoids interference from cross-year data |
Three common configuration mistakes
- When creating a document collection, PDF parsing parameters are not specified, resulting in department final account statements in PDF format failing to be parsed correctly. This occurs because corresponding parsing rules are not configured in API requests, and the default parsing logic does not adapt to the nested table structure of financial PDF reports.
- Statistical scope deviations appear in knowledge base retrieval results. Investigation finds that an incompatible vector model is used. This occurs because the adapted vector model is not switched based on the characteristics of financial structured data, leading to inaccurate field semantic recognition.
- An error prompting "field format mismatch" or empty results is returned during retrieval. This occurs because field mapping rules for data from multiple systems are not unified, leading to failure to recognize standardized financial fields during retrieval.
How to confirm the configuration is correct
- Upload a single annual final account PDF document, call the API to create a collection, and check if the parsing result fully extracts tables and explanatory text.
- Initiate a query for "202X annual income and expenditure final account differences", and check if the returned results include the corresponding year’s final account explanations and report data.
- View the retrieval configuration panel, confirm that the
similaritythreshold,limitentry count, andusingReRankswitch are all enabled per the set values. - After switching the vector model, compare the semantic matching degree of two retrieval results, confirm that the model switch takes effect.
The values provided 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.