What data looks like for this category
This category’s data comes from investment supervision systems of the state-owned assets supervision and administration commission, enterprise ERP and project management systems, industry market data, and audit and financial statements. Commonly processed documents include annual investment plans, project proposals, feasibility study reports, due diligence reports, asset appraisal reports, and post-investment evaluation reports. The data update cycle is quarterly. Most documents include fields such as project number, investment amount, implementation period, and risk control review opinions. The length of individual documents varies significantly.
What constraints do these characteristics impose on knowledge base retrieval and recall
Multi-source heterogeneous data sources create differences in document formats and field standards. The retrieval process must support unified parsing and indexing of cross-system data, to avoid missed recall caused by format incompatibility. The quarterly update rhythm requires the knowledge base synchronization strategy to align with the business cycle, ensuring review uses valid data from the latest quarter. The wide variation in document length requires retrieval configuration to balance reasonable long-text chunking and recall accuracy, preventing semantic breaks from over-splitting long documents. The review link is tied to state-owned asset security, so retrieval results must cover core documents across the full project lifecycle, meeting requirements for investment decision-making committee materials and risk control verification.
How to set configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the density of technical terminology in documents of this category, ensuring the relevance of recall results to declaration materials meets review requirements |
limit | Top 8 entries | Balances retrieval efficiency and information completeness, covering multi-dimensional documents required for project approval and post-evaluation |
searchMode | Semantic retrieval | Adapts to the professional text focus of most content in this category, avoiding ambiguity issues with keyword retrieval |
usingReRank | Enabled | Improves recall sorting accuracy in long document scenarios, filtering low-relevance chunk results |
datasetSearchUsingExtensionQuery | Enabled | Supplements retrieval expansion for fields such as project number and investment amount, covering multi-dimensional query requirements |
collectionIds | Configured by business domain | Distinguishes knowledge base collections for approval documents and post-evaluation documents, accurately matching the retrieval scope for different review links |
Three common misconfigurations
- Phenomenon: Retrieval results include documents outside the target business domain, such as mixing industry market data into project approval declaration material review results. Cause: The
collectionIdsretrieval scope is not configured, leading to cross-domain data recall. - Phenomenon: When creating a general knowledge base in a local deployment version, the interface does not display an image understanding model option, with a prompt that the model is not loaded. Cause: Environment dependencies for the image understanding model are not configured per deployment guidelines, and the corresponding service process is not started.
- Phenomenon: Retrieval results cannot be filtered by business fields such as project number and investment amount, only returning full relevant documents. Cause: Custom filter fields are not enabled in the knowledge base configuration, and field mapping configuration with the business system is not completed.
How to confirm configurations are correct
- Submit a retrieval request for a single project proposal, verify that returned document types belong to the preset approval review business domain, and confirm the
collectionIdsconfiguration takes effect. - Adjust the
similaritythreshold and run multiple retrievals, observe changes in result count and relevance, and confirm the threshold setting meets business review accuracy requirements. - After enabling the
usingReRankfunction, compare retrieval result sorting before and after enabling, confirm the reranking function operates normally, and improve display priority for highly relevant results. - Submit a query request containing a project number, verify that retrieval results only return documents for the corresponding project, and confirm custom filter field configuration is correct.
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.