What this category of data looks like
Business processing data for new energy development primarily comes from government project filing systems, grid connection approval platforms, construction permit archives, and project feasibility study reports. Data updates are triggered by approval nodes, including key stages such as project filing, construction permits, and grid connection acceptance. There is no fixed batch update cycle. Document structures include fields such as project number, filing number, land scope, installed capacity, approval progress, and handler information. Units include professional measurement standards such as kilowatts, hectares, and workdays. Individual document lengths range from hundreds to tens of thousands of characters.
What constraints these characteristics impose on the vector models and indexing link
Professional fields and units in new energy development business require vector models to accurately identify technical terms, and avoid term splitting caused by overly small chunking. Non-fixed update cycles require indexing systems to support incremental triggers, without full reconstruction. Long-text feasibility study reports and multi-field documents require vector indexes to adapt to input content of different lengths. They also need to avoid matching errors caused by similar expressions across project approval nodes. Repeated approval circulation copy files also increase configuration requirements for index deduplication.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the field length of new energy development project documents, avoids splitting technical terms and units of measurement |
| Overlap Length | 100–150 characters | Preserves approval process context across chunks, prevents key steps from being truncated |
| maxToken | 4096 | Matches the input limit of mainstream open-source embedding models, adapts to long-text feasibility study report content |
| Similarity Threshold | 0.72–0.78 | Distinguishes approval node expressions for similar projects, avoids cross-project matching errors |
| Single Search Limit | Top 6 entries | Covers key approval stages across the entire new energy development process, prevents omission of core information |
| defaultToken | 1024 | Configures the context length for basic retrieval, adapts to routine business consultation scenarios |
Three common configuration errors
- Phenomenon: When uploading a document, it displays 8 chunks, but changes to 13 chunks after completion, with duplicate indexes. Cause: Document deduplication configuration is not enabled, and copy files generated during approval circulation are repeatedly identified as new content.
- Phenomenon: After adding ollama qwen2.5 and bge-m3 models, the corresponding options do not appear in the text understanding model dropdown menu when creating a knowledge base. Cause: The embedding model is not bound to the vector model configuration group of the knowledge base, or the model deployment port is not open.
- Phenomenon: Batch retraining of vector models is not supported, and only single-file parameter adjustment and upload are available. Cause: Batch index update function is not enabled, and trigger rules for incremental update are not configured.
How to verify correct configuration
- Upload a standard new energy project approval document, and check whether the chunk results match the preset chunk length and overlap length.
- Test retrieval of different approval node content for the same project, and confirm that returned results do not contain cross-project matching errors.
- Try batch uploading multiple similar documents, and check whether the number of segments after index completion matches expectations.
- View the model management page, and confirm that the embedding model is associated with the vector configuration group of the current knowledge base.
The values listed above are common starting points, and should be evaluated against your 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.