What the data for this category looks like
Data for this category primarily comes from the budget management integrated system, department staff establishment system, prior year final accounts, project library, and Ministry of Finance budget preparation notices. The update frequency is annual. Commonly processed documents include department budget drafts, project application forms, performance target forms, staff expense calculation forms, and budget approvals. Fields include staff establishment count, project expense amount, performance target indicators, and budget approval amount. Common units are ten thousand yuan, people, and items. The data structure is adjusted slightly according to budget year policies.
What constraints these characteristics impose on the vector models and indexes stage
Reference only; verify against the target FastGPT release and deployment documentation.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Department budget documents are mostly structured long texts. Excessively long chunks will lose semantic connections, while excessively short chunks will destroy the complete logic of projects and expenses. This range is determined based on actual testing. |
| Overlap Length | 100–150 characters | Prevents semantic breaks across projects after chunking, ensuring that the front-back association of budget projects is covered by vector embeddings. |
| Similarity Threshold | 0.6–0.8 | Department budget documents have high professionality. Semantic matching requires strict filtering of low-relevance results to avoid non-budget data interfering with consultation replies. |
| Single Search Limit | Top 5 results | Budget consultations mostly require precise matching of projects or clauses. Too many results increase the cost of reply screening. |
maxToken | 8192 | Adapts to long text processing for department budget documents, avoiding truncation of key content during vector encoding. |
defaultToken | Configured exclusively for the knowledge base | Matches the field encoding requirements of department budget data, ensuring consistent vector dimensions of documents from different data sources. |
Three common mistakes
- Knowledge base imported datasets remain in the indexing status for a long time. Phenomenon: After uploading department budget-related documents, the interface status stays as indexing with no progress updates. Cause: No batch indexing task matching the annual data update frequency is configured, or the chunk length setting exceeds the system's processing limit, causing the indexing process to block.
- Semantic similarity values are abnormally high after switching vector models. Phenomenon: After using a specified model, similarity results show values above 10000, and search filtering cannot be set normally. Cause: The system's built-in similarity normalization configuration is not adjusted, resulting in unstandardized raw model output that exceeds the preset threshold range.
- Knowledge base fails to load model channels normally. Phenomenon: When configuring an embedding model, no available channels are displayed, and the corresponding model cannot be selected. Cause: The API key or service address of the corresponding model is not added in the system configuration, or the channel permission is not opened to the currently used knowledge base scope.
How to confirm the configuration is correct
- Upload a single department budget draft document, check if the chunked fragments retain complete core fields such as project name and expense amount, with no forced truncation.
- Input a typical budget consultation question, such as "What are the performance target requirements for a certain project", check if the relevance of returned results meets business expectations, with no irrelevant final account or expenditure data.
- Check the indexing task logs, confirm that all imported budget documents have completed vector encoding, with no error fields or unprocessed data sources.
- Adjust the similarity threshold configuration item, verify whether the change in the number of search results under different thresholds meets business screening requirements.
The values listed above are common starting points, and should be evaluated against the reader's own specific 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.