What this category of data entails
The core documents of this category include performance target forms, self-assessment forms, performance evaluation reports, supporting material lists, and rectification reports. Data mainly comes from the unified budget management system, project application materials, fund payment data, business system ledgers, and satisfaction surveys. The data update frequency is annual. Each single document usually contains fields such as project number, fund amount, performance indicators, and supporting material links. Units include ten thousand yuan, person-times, completion rate, and others. Most document structures combine structured tables and paragraph descriptions.
What constraints these characteristics bring to the knowledge base retrieval and recall link
Annual update frequency means the knowledge base does not need frequent incremental synchronization. Only full retraining is required once per year. The mixed structure of structured tables and paragraph descriptions requires retrieval to support content recall for mixed formats. Multiple data sources require retrieval to support cross-system associated keyword matching. This accurately locates the correspondence between project outputs, fund expenditures, and performance targets. The accuracy of retrieval results directly affects the compliance of performance evaluation reports. Irrelevant content must be strictly filtered to avoid interfering with subsequent document drafting.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the accuracy requirements of fiscal performance documents, and avoids recalling irrelevant non-performance materials |
limit | Top 10 | Covers multiple types of documents required for performance evaluation, such as supporting materials and historical reports, while controlling context length |
searchMode | vector + keyword | Adapts to mixed-structure performance documents, balancing semantic matching and precise keyword matching |
usingReRank | Enabled | Performs secondary ranking on recall results, improves the ranking of highly relevant documents, and adapts to the precise matching requirements of performance indicators |
datasetSearchUsingExtensionQuery | Enabled | Expands relevant performance terms based on the original question, covering synonyms and industry expressions, and improves recall comprehensiveness |
collectionIds | Specify the exclusive dataset ID for performance evaluation | Isolates non-performance documents to avoid irrelevant content being mixed into retrieval results |
Three Common Misconfigurations
- After performing a full retraining of the knowledge base, some historical performance reports fail to load correctly. The symptom is that prior-year performance self-assessment forms cannot be retrieved after training completes. The cause is that full import is conducted without clearing the original dataset, leading to conflicts between new and old data or failure to cover some documents.
- The number of retrieval results is far lower than expected, with only a small amount of content returned. The symptom is that searching for "project fund use performance" returns no more than 3 results. The cause is that the
similaritythreshold is set too high, filtering out relevant documents that meet requirements, or thelimitvalue is set too low to cover all required content. - The exported knowledge base file size is only hundreds of KB, and the content is incomplete. The symptom is that the exported file only contains a small amount of metadata. The cause is that only part of the dataset content is exported, not all associated supporting material files are selected, or the full export mode is not chosen.
How to Confirm the Configuration Is Correct
- Perform a test retrieval for performance indicators, check the relevance of recall results, and adjust configurations such as
similarityandusingReRankuntil the expected outcome is achieved. - Import a test Markdown file containing image links, replace local image links with online image bed addresses, and verify the loading status to confirm that the image processing configuration takes effect.
- After performing a full retraining, check that the total number of documents in the dataset matches the total number of imported original files, confirming that no training omissions occurred.
- Export the knowledge base file, check that the exported content size matches the storage capacity of the original data, confirming that the export scope covers all documents.
The values provided here 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.