What this category of data looks like
This category covers end-to-end investment process documents from central enterprise strategic investment departments, project companies, state-owned assets supervision departments, and external consulting and evaluation institutions. Documents include annual investment plans, project proposals, feasibility study reports, due diligence reports, asset evaluation reports, and investment post-evaluation reports. Data comes primarily from state-owned assets supervision and administration investment systems, enterprise ERP and project management systems, industry market data, and audit financial statements. It updates quarterly. Most documents mix structured and semi-structured content. They include fixed fields such as project number, investment amount, implementation cycle, and risk control opinions. No unified unit requirements apply. Matching the original format of each document is required.
Constraints on knowledge base retrieval and recall
Quarterly updated data requires the knowledge base to support incremental synchronization and full re-updates. This prevents outdated historical data from being recalled. The mixed structure of multiple document types requires the retrieval link to distinguish field attributes across different documents. It must prioritize recalling feasibility study reports and due diligence content closely tied to investment approval and post-evaluation work. Fixed fields in the data require retrieval to support precise matching for items such as project number and investment amount. It must also support unstructured text such as risk control opinions. High compliance requirements in state-owned assets supervision scenarios require retrieval results to trace back to original data sources. It must also support filtering recall ranges by data update time.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Meets the strictness requirements for state-owned enterprise project documents, filters low-correlation non-investment content |
limit | Top 8 entries | Balances information density for ledger summary and context length for report generation. Prevents exceeding model processing limits |
searchMode | Paragraph-first | Adapts to the structure of documents in this category, which are mostly long-paragraph feasibility study and post-evaluation reports. Preserves complete business logic |
usingReRank | Enabled | Optimizes recall ranking in multi-document mixed scenarios. Prioritizes content related to core indicators of approval and post-evaluation work |
datasetSearchUsingExtensionQuery | Enabled | Supplements extended retrieval for fields such as project number and investment cycle. Improves precise matching accuracy |
Three Common Mistakes
- Issue: The share of old documents in knowledge base recall results is too high. The recall performance of newly deployed embedding models does not meet expectations. Cause: Full re-embedding of the entire dataset was not completed. Only incremental documents were updated. This leads to inconsistent vector spaces between new and old models.
- Issue: An error is returned for passed parameters when calling the interface to create a text collection. Cause: Required fields such as
collectionIdswere not submitted in accordance with FastGPT v4.17.0 interface specifications. Or the parameter format does not meet requirements. - Issue: Document content uploaded via Feishu knowledge base synchronization cannot be retrieved after synchronization. Cause: Access permissions for Feishu documents were not granted to the FastGPT synchronization account. Or the file format was not recognized by the supported parsing engine.
How to Verify Proper Configuration
- Run a precise retrieval for a project number. Check if returned results include core content of the corresponding document. Adjust the
similaritythreshold to a range that meets business strictness requirements. - View the source field of retrieval results. Confirm that it matches the preset data source range. Verify that the
collectionIdsconfiguration takes effect. - Compare recall sorting results with and without
usingReRankenabled. Confirm that the reranking function optimizes result order according to business needs. - Upload a new document updated quarterly. Run a retrieval test. Confirm that the new data can be recalled normally. Verify that the incremental synchronization configuration works correctly.
The values listed are common starting points. They 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.