Reference only; verify against the target FastGPT release and deployment documentation.
What this category of data looks like
This category of data mainly comes from the national land spatial planning one-map, project approval and supervision platform, land use status database, and ecological protection redline data. Daily documents include land pre-examination and site selection opinions, construction land planning permits, construction engineering planning permits, redline drawings, and planning conditions, covering fields such as project site selection scope, land scale, control requirements, and redline coordinates. Data updates are irregular with no fixed cycle, and each individual document contains both structured approval information and non-spatial text descriptions.
Constraints Imposed on Vector Models and Indexes
This category of data contains spatial coordinate content and structured approval fields. When chunking, avoid splitting associated spatial scope and approval requirements to prevent the vector index from losing key associated information. Data sources are scattered and updates are irregular, so index logic that supports incremental updates must be configured to avoid resource consumption caused by full reconstruction. Each individual document contains both text descriptions and numerical indicators. The vector model must adapt to the encoding of mixed-format content to ensure that semantic associations between different types of fields are accurately captured.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | This category of documents contains spatial descriptions and approval text. This length preserves the integrity of a single approval logic |
| Overlap Length | 100–150 characters | Prevents spatial coordinates and associated text from being split after chunking, ensuring complete associated information |
defaultToken | 1024 | Adapts to the default token setting of basic embedding models, avoiding initial retrieval exceptions |
maxToken | 8192 | Adapts to the maximum token limit of mainstream embedding models, avoiding content truncation |
| Similarity Threshold | 0.72–0.80 | Calibrated based on actual measurements, distinguishing similarity and compliance differences in project site selection |
| Single Search Limit | Top 6 entries | Approval for this category requires checking multiple precondition documents, and an appropriate number of results can cover associated information |
Three Common Configuration Mistakes
- Symptom: Search response times out after the knowledge base index is completed, taking more than 30 seconds. Cause: The chunking strategy was not adjusted for the large-volume spatial documents of this category, leading to token overload in single-batch vector retrieval.
- Symptom: Knowledge base calls still return embedding model errors after configuring
text-embedding-ada-002. Cause: The embedding model was not correctly associated with FastGPT's API channel, or port mapping of the model container was not completed during local deployment. - Symptom: A large number of irrelevant redline drawing data appears in search results. Cause: The similarity threshold was set too low, failing to filter non-associated spatial documents, leading to insufficient retrieval accuracy.
How to Verify Correct Configuration
- Upload a single land pre-examination opinion document, run a chunking test, and confirm that the chunking results do not split spatial coordinates and approval requirements.
- Call the embedding model interface, verify that the returned vector dimensions match the configured parameters, and no truncation errors occur.
- Simulate a project approval query, check that the search results include the corresponding precondition documents, and the number of results aligns with the single search limit setting.
- Upload an updated planning condition document, run an incremental index test, and confirm that the new data can be retrieved normally.
The values provided 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.