What This Category of Data Entails
Daily documents for this category include territorial spatial master plans, detailed plans, special plans, planning implementation evaluation reports, and public opinion summaries. Data primarily comes from the territorial spatial basic information platform, natural resource survey and monitoring data, population and economic statistics, and remote sensing images. Update frequency is once per year. Most documents contain spatial coordinate fields, planning indicator parameters, and departmental opinion text. Some files have spatial layer attachments. Overall data volume increases with planning hierarchy, and content uses multi-source heterogeneous formats.
Constraints on Vector Models and Indexing
The multi-source heterogeneous data characteristics of this category require the indexing link to support both text and spatial attribute data formats. The annual update rhythm means high-frequency incremental indexing is unnecessary. Single full indexing involves a large data volume, so the system must support batch processing capabilities. Mixed storage of spatial coordinate fields and planning indicator parameters requires vector models to balance text semantic matching and structured field association. The coexistence of short public opinion text and long planning report text requires unified chunking rules to avoid semantic fragmentation. Indexing result accuracy directly affects subsequent planning approval workflows, so the false match rate must be strictly controlled.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the paragraph length of planning reports and the text length of public opinions, avoids semantic truncation |
| Overlap Length | 100–150 characters | Retains cross-chunk spatial indicator association information, prevents loss of key constraints at chunk boundaries |
| Similarity Threshold | 0.72–0.85, calibrated via actual testing | Distinguishes similar approval entries in planning text from irrelevant reference materials, reduces false match probability |
| Single Index Limit | 500 documents | Matches the annual planning document batch processing rhythm of natural resource departments, avoids single-batch indexing timeouts |
defaultToken | 4096 | Adapts to the input length limit of mainstream open-source embedding models, supports long paragraph chunking of planning reports |
| Rerank Switch | Enabled | Optimizes result ranking after multi-source data retrieval, improves matching accuracy for content related to planning approval |
Three Common Mistakes
- Indexing tasks remain in processing for a long time, with timeout errors returned in logs. The cause is failure to adjust the single index limit for annual updated full data; loading too many documents at once exhausts system resources.
- The text understanding model dropdown menu does not show the added ollama qwen2.5 model when creating a knowledge base. The cause is failure to configure the adaptation relationship between the embedding model and text understanding model, or the model deployment port is not opened to an accessible range.
- The number of matching results returned when searching planning files is far lower than expected. The cause is that the chunk length is set too large, leading to long text being truncated so that key approval conditions cannot be matched.
How to Verify Proper Configuration
- Submit 10 typical planning documents for test indexing, check whether the task completion time meets business expectations.
- Retrieve indexed public opinion summary text, confirm that the returned results contain core opinion keywords.
- Check the model dropdown list, confirm that the added ollama qwen2.5 and bge-m3 models are displayed in the selectable range.
- View indexing logs, confirm that there are no error messages about spatial field parsing failures or format incompatibility.
The values listed above 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.