What this category of data looks like
Data for this category primarily comes from the budget management integrated system, project declaration materials, fund payment data, business system ledgers, and satisfaction surveys. Commonly processed documents include performance target forms, self-assessment forms, performance evaluation reports, supporting material lists, and rectification reports. Most document structures include project association identifiers, performance indicator entries, fund usage association items, evaluation result fields, and rectification issue records. The data update cycle is annual, with each single update covering all completed performance-related project data for the full year.
Constraints Imposed on Knowledge Base Retrieval and Recall Links
The annual update feature means the knowledge base does not require frequent incremental synchronization, and a full refresh mode can be used to reduce resource usage from real-time synchronization. The coexistence of multiple document types requires the retrieval system to adapt to a mixed splitting logic for structured tables and unstructured reports, avoiding breaking the association between performance indicators and supporting materials during splitting. Multiple data sources have field differences, so unified metadata mapping rules are needed to ensure that information for associated projects can be fully recalled. Single annual data covers all projects, so a single retrieval may return a large number of results. Recall scope must be limited to avoid context overload affecting subsequent generation.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches semantic similarity for performance-related text, prevents irrelevant results from being included if set too low, and avoids missing relevant supporting materials if set too high |
limit | Top 8–12 results | Sets the upper limit for recall volume for annual data, prevents excessive results from causing context overflow, and aligns with the information density needs of performance evaluation consultations |
searchMode | hybrid | Adapts to mixed retrieval needs for structured tables and unstructured reports, matching both keyword and semantic features |
usingReRank | Enabled | Reorders recall results by relevance, optimizing the sorting effect of performance evaluation reports and their associated supporting materials |
datasetSearchUsingExtensionQuery | Enabled | Generates extended queries based on user questions, covering semantic coverage for specific scenarios such as performance indicators and rectification matters |
collectionIds | Bind to the exclusive dataset ID for performance evaluation | Restrict retrieval scope to documents related to fiscal expenditure performance evaluation, preventing mixing with data from other fiscal categories |
Three Common Configuration Mistakes
- Symptom: Retrieval logs show response time exceeding 10 seconds, with noticeable front-end conversation lag. Cause: Failed to bind exclusive
collectionIds, retrieval scope covers unrelated fiscal documents, leading to excessive recall volume. - Symptom: Number of retrieved results exceeds 15, triggering a context overflow error during AI response generation, causing the answer to stop midway. Cause: Did not set a reasonable
limitparameter, excessive recalled content exceeds the preset context window limit. - Symptom: Retrieved performance indicator chunks do not carry the supporting material metadata for their corresponding project, with associated information missing from output results. Cause: Did not enable
datasetSearchUsingExtensionQuery, failed to bind and pass metadata with retrieved chunks, leading to information silos.
How to Verify Proper Configuration
- View retrieval logs in the FastGPT backend, confirm only performance-related documents from the bound exclusive dataset are returned, with no cross-category results included.
- Initiate a query that links performance indicators and supporting materials, check that returned results include metadata fields and are bound to the retrieved chunk content.
- Simulate a multi-result query scenario, confirm the number of returned results falls within the preset
limitrange, with no excess results. - Test a standard retrieval request, confirm response time meets business expectations with no noticeable lag.
The values given 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.