What the data for this category looks like
The core data for this category comes from the budget management integrated system, centralized treasury payment system, accounting system, and asset management system. The update frequency is annual. Daily documents include department final account statements, balance sheets, income and expenditure final account statements, final account explanations, and government financial reports. Data fields are fixed and must strictly match fiscal standards, with no temporary new fields added. Data is submitted by entities such as budget unit finance departments, treasury departments of finance bureaus, and competent department finance divisions. It must go through multi-level review before centralized submission, and the single submission data volume is large.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The annual update feature means the knowledge base does not require frequent refreshes, and data can be imported in annual batches. However, the single imported dataset scale is large, so reasonable chunking rules must be set to avoid parsing timeouts. The unified caliber requirement from multiple system sources means fixed fields such as subject codes and report numbers must be strictly matched during retrieval, to avoid caliber deviation caused by semantic expansion. Multiple document types require differentiated retrieval weights for statement-type and explanation-type documents, to ensure core statement data is recalled first. Retrieval for this category focuses on actual revenue and expenditure, assets and liabilities, and difference explanations, which differs from the retrieval focus of the budget preparation link. Therefore, the search mode must adapt to exact matching of fixed fields. The annual centralized submission feature requires the retrieval link to quickly locate historical data for a specific year, to avoid interference from cross-period data.
Reference only; verify against the target FastGPT release and deployment documentation.
How to set the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
Similarity threshold (similarity) | 0.75–0.85 | Final account data has strict caliber requirements, need to filter low-match irrelevant results to ensure retrieval results meet fiscal review requirements and avoid issues like inconsistent book and actual records |
Single search limit (limit) | Top 5–8 results | Annual report data volume is large, too many retrieved results increase context processing burden, while avoiding missing core matching items |
Search mode (searchMode) | Exact match mode | Match fixed fields such as subject codes and report numbers to avoid caliber deviation from semantic expansion, adapting to the business retrieval focus of this category |
Rerank toggle (usingReRank) | Enabled | Multiple systems have many similar fields, reranking can improve the sorting priority of core statement data to ensure reviewers view key content first |
Extended query toggle (datasetSearchUsingExtensionQuery) | Disabled | Final account queries require strict matching of original caliber, extended queries may introduce irrelevant semantics and lead to review deviation |
Dataset binding ID (collectionIds) | Bind fiscal system datasets for the corresponding year | Isolate data by year to avoid cross-period data confusion, complying with the time range requirements of final account review |
Three Common Mistakes
- Issue: After calling the API to create a knowledge base, only the final ready or failed status is displayed, with no real-time parsing status available. Cause: The dataset status query API was not called correctly, or the API permission was not configured to return parsing progress in real time.
- Issue: After switching the vector model, the semantic similarity return value exceeds 10000, and the search filtering function fails. Cause: A similarity parameter outside the 0-1 range was used incorrectly, or the vector model output format does not match the system requirements.
- Issue: After configuring the workflow, the knowledge base search link terminates directly during execution. Cause: The retrieved results exceed the context length limit, or the bound
collectionIdsdo not include datasets from the corresponding business systems, resulting in no matching results.
How to Confirm the Configuration Is Complete
- Check that the bound
collectionIdsonly include datasets for the corresponding year from the budget management integrated system, centralized treasury payment system, and other systems, to avoid mixing irrelevant data. - Retrieve a specific subject code or report number to verify that the
similaritythreshold filters out results with matching degrees below the set value, ensuring retrieval accuracy meets fiscal review requirements. - Call the API to query the knowledge base parsing status, confirm that clear status fields of parsing, ready, or failed are returned, with no null values or abnormal status codes.
- After enabling
usingReRank, retrieve a query containing multiple similar fields to verify that the sorting priority of core statement data is higher than that of auxiliary explanation documents.
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.