What the data for this category looks like
Data related to department budget preparation typically originates from the unified budget management system, department staff establishment system, prior-year final accounts, project library, and Ministry of Finance budget preparation notices. Update frequency is annual. Daily documents include department budget drafts, project application forms, performance target forms, staff expense calculation forms, and budget approvals. Document fields include budget subject codes, staff establishment numbers, project performance indicators, expense calculation calibers, and more. Common units are ten thousand yuan, people, number of projects, and similar. Individual document text lengths vary widely, from notices of a few hundred characters to summary calculation forms with tens of thousands of characters.
What constraints these characteristics impose on knowledge base retrieval and recall
Department budget preparation data has three key traits: annual updates, scattered sources, and large differences in document structure. These traits create multiple constraints for knowledge base retrieval and recall. Annual updates require the knowledge base to support incremental data synchronization by annual batches. This avoids redundant or missed invalid content across cycles. Documents from multiple systems have inconsistent formats. Field alignment and format standardization must be completed first to ensure consistent recalled content. Documents include precise fields such as budget subject codes and staff establishment numbers. Retrieval must balance semantic matching and precise field matching to prevent irrelevant results from being included. Individual document text lengths vary widely. The scale of single-recall content must be controlled to avoid context overflow that reduces answer accuracy.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Department budget-related documents have long text and precise fields. A threshold that is too low will introduce irrelevant historical data, while a threshold that is too high will miss relevant calculation calibers or notice content. This interval is determined after actual testing. |
limit | Top 8 entries | Department budget preparation-related documents are mostly a mix of structured and semi-structured formats. Recalling 8 entries in a single query can cover reference content for typical consultation scenarios while controlling context length. |
searchMode | Full-text retrieval | Adapts to the mixed structure of multiple types of documents, can match both text semantics and field keywords, and meets scenarios such as caliber queries and project application requirements in budget preparation. |
usingReRank | Enabled | A large number of similar subject codes or calculation rules exist in budget-related documents. Reranking can filter out entries most relevant to the consultation question, improving answer accuracy. |
datasetSearchUsingExtensionQuery | Enabled | Budget preparation consultations often involve term expansion. For example, "staff expenses" can be expanded to "staff salaries, social insurance, public expenses", optimizing the recall scope. |
collectionIds | Bind the budget document collection for the corresponding year | Binding collections by year prevents cross-year data from being mixed in, ensuring recalled content matches the current preparation cycle. |
Three common errors
- Phenomenon: After uploading a department budget calculation form, the system displays 12 segments when the upload finishes. Refreshing the page later changes the count to 17, with duplicate indexed entries. Cause: The document contains duplicate paragraphs with budget subject codes, and no clear paragraph boundary was set during chunking. This causes the system to repeatedly split the same content.
- Phenomenon: When retrieving budget preparation calibers, returned results only list citation entries at the bottom. The answer text does not link to the corresponding citation content. Cause: The binding logic between answers and citations was not configured, leading to a disconnect between citation entries and answer text.
- Phenomenon: When retrieving a collection with a large number of long documents, the system prompts a retrieval timeout. Cause: The maximum document loading length for single retrieval was not limited. The excess of long budget documents causes retrieval time to exceed the threshold.
How to confirm configurations are properly set
- Upload a single department budget draft document. Check the number of chunks after splitting, confirm there are no duplicate split identical paragraphs.
- Enter the consultation question "staff expense calculation caliber". Verify that the citation entries of the returned results match the answer content, confirming the citation logic works correctly.
- Switch to bind document collections for different years. Confirm that retrieval results only include budget-related documents for the corresponding year, with no cross-year data included.
- Adjust the
similarityparameter to 0.75. Test the relevance of retrieval results, confirm that no excessive irrelevant content is included.
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.