Reference only; verify against the target FastGPT release and deployment documentation.
What the data for this category looks like
Core documents for this category include territorial spatial master plans, detailed plans, special plans, planning implementation evaluation reports, and public opinion summaries. Data is mainly sourced from the territorial spatial basic information platform, natural resource survey and monitoring data, population and economic statistical data, and remote sensing images. The data update frequency is annual. Most documents contain fields such as spatial coordinates, constraint indicators, and regional layout plans. Units include hectares, square kilometers, population counts, and others. Total data volume increases as the planning area expands.
What constraints do these characteristics impose on knowledge base retrieval and recall
Annually updated data sources require the knowledge base to run full synchronization regularly, to avoid recalling expired planning content. Format differences across multi-source heterogeneous data (spatial data, statistical data, image data) require retrieval logic to adapt to different types of text and structured data, to prevent matching failures for unstructured spatial data. Documents contain a large number of professional fields and spatial coordinates, so retrieval must prioritize matching core planning constraint items. Short text data such as public opinion summaries must use different recall weights from large planning reports, to avoid low-value content interfering with core retrieval results.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category’s documents are mostly professional planning texts. This threshold filters low-match irrelevant content and avoids recalling unrelated planning clauses |
limit | top 8–12 entries | Planning documents have long content. Too many recalled entries will exceed the context window, while too few will fail to cover multiple constraint conditions |
usingReRank | Enabled | When retrieving mixed multi-source heterogeneous data, re-ranking can prioritize matching text blocks related to spatial constraints and core indicators |
datasetSearchUsingExtensionQuery | Enabled | Planning questions often contain professional terms. Extended queries supplement synonym matching and improve recall rates for professional content |
collectionIds | Specify the folder ID of the planning knowledge base | This category’s data is independent of other natural resource data. Limiting the scope avoids interference from unrelated knowledge base content |
searchMode | Hybrid retrieval of vector + keyword | Spatial planning documents have both structured indicators and unstructured descriptions. Hybrid mode balances exact matching and semantic matching |
Three common mistakes
- Phenomenon: Retrieval fails to limit searches only to the specified planning knowledge base, and natural resource data from other categories is recalled. Cause: The
collectionIdsparameter is not configured correctly, and the target knowledge base folder ID is not specified. - Phenomenon: Retrieval results do not return the preset question-and-answer pair text from the knowledge base, and generalized generated content appears. Cause: Question-and-answer pairs are not separately collected into a dedicated knowledge base, or preset question-and-answer library content is not prioritized for matching.
- Phenomenon: Text blocks are lost when retrieving long planning reports, and recalled content is incomplete. Cause: The
similarityparameter is set too high, filtering some relevant content, or thelimitparameter is set too low, failing to recall a sufficient number of text blocks.
How to confirm the configuration is correct
- Enter the FastGPT knowledge base retrieval test interface, enter professional questions related to territorial spatial planning, and check if retrieval results only include target planning documents and preset question-and-answer pairs.
- Adjust the
similarityparameter, observe changes in the matching degree of retrieval results, and confirm the threshold meets the matching requirements for professional texts. - After enabling
usingReRank, compare the sorting of retrieval results before and after enabling to confirm the re-ranking logic takes effect. - Submit fully updated annual data, check that the vectorization and retrieval processes run normally, with no over-limit errors.
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.