What This Category of Data Looks Like
Data for this category comes from the unified budget management system, project declaration materials, fund payment data, business system ledgers, and satisfaction surveys. The update frequency is annual. Commonly processed documents include performance target forms, self-assessment forms, performance evaluation reports, supporting material lists, and rectification reports. Most fields relate to project numbers, fund amounts, performance indicators, supporting material numbers, and similar content. There are no unified fixed units, so processing must match the native fields of the documents.
Constraints Imposed by These Characteristics on the Knowledge Base Retrieval and Recall Link
The annual update frequency means the knowledge base does not require high-frequency synchronization, and full updates can be set for fixed cycles. The mixed structure of multiple document types requires retrieval to support filtering recall results by document type. Fields involve highly professional performance-related content, so precise recall matching native fields is required to avoid generalized matching interfering with business judgments. Supporting material lists are closely linked to performance reports, so cross-document associated recall must be supported to ensure retrieval results cover the complete performance supporting evidence chain.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Documents in this category have highly professional fields, so low-match irrelevant content must be filtered to avoid interfering with performance indicator verification |
limit | Top 8–12 results | Performance evaluation needs to cover multiple types of documents including targets, self-assessments, supporting materials, and similar content. An appropriate number of recall results ensures coverage of complete business processes |
searchMode | vector | Data in this category is a mix of structured and semi-structured documents, and vector retrieval can accurately match professional fields |
usingReRank | Enabled | Recall results need to be reordered by business relevance to ensure core performance indicators and supporting materials are displayed first |
datasetSearchUsingExtensionQuery | Enabled | Extended queries can be used to supplement professional terminology, improving recall accuracy for performance evaluation-specific vocabulary |
collectionIds | Configure corresponding collection IDs by document type | Documents in this category have clear classification, and filtering by collection reduces interference from irrelevant data |
Three Common Misconfigurations
- Imported performance evaluation report PDFs cannot be opened in retrieval results. The cause is that external link jump permissions for knowledge base files are not configured, or documents are not uploaded to a storage path that supports direct access.
- After enabling
usingReRank, video memory remains occupied after retrieval completes. The cause is that the logic to automatically release reranking model resources after retrieval is not set, resulting in model resources not being recycled. - An Invalid array length error occurs after selecting the question answering splitting function. The cause is that empty fields or ultra-long text blocks are not filtered during splitting, causing abnormal array parameters.
How to Confirm Proper Configuration
- Upload a performance target form document to the corresponding knowledge base collection, initiate a query containing performance indicator keywords, and verify that valid fragments of the document are included in the recall results.
- Check system video memory usage, confirm that the video memory occupied by the reranking model is automatically released after retrieval is initiated and results are obtained.
- Test the question answering splitting function, input the complete text of a self-assessment form, and confirm that no Invalid array length error is reported.
- After configuring the
similaritythreshold, initiate a query with low matching accuracy, confirm that results are filtered, and only content meeting the threshold is retained.
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.