What this category’s data looks like
Survey and design data has four primary sources: project survey reports, official design drawing descriptions, industry standard atlases, and project approval documents. Update frequency aligns with project progress. Updates are applied on demand during a single project cycle, and synced to the matching knowledge base content when industry standards are revised. Document structures include long text details, structured fields (such as project number, survey point coordinates, compliance clause number), and quantified data with units (like meters, square meters, cubic meters). Individual document content can span a wide range; some drawing description documents reach tens of thousands of characters.
Constraints imposed on knowledge base retrieval and recall
Long text documents with wide content spans can break contextual connections after chunking. Adjust chunking and overlap parameters to preserve complete logical flow. Structured fields and quantified units require retrieval to balance precise matching and semantic understanding. This prevents irrelevant results from generalized retrieval. Survey and design data across different projects is independent and has significant version differences. Strictly limit retrieval scope to the corresponding project dataset. Frequent updates to industry standards require the knowledge base synchronization mechanism to adapt to rapid content changes.
Configuration settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Compliance clauses and project requirements for survey and design need high matching accuracy to avoid irrelevant results disrupting business guidance |
limit | Top 8–12 items | Single survey reports have many chunks. Excessive recall results exceed the context window and impair process guidance logic |
searchMode | Precise matching + semantic retrieval | Balances retrieval needs for structured fields (such as standard number, project number) and semantic survey steps |
usingReRank | Enabled | Redundant or low-correlation results appear after long text chunking. Re-ranking improves the accuracy of recalled results |
datasetSearchUsingExtensionQuery | Disabled | Business process guidance requires strict matching of original questions. Extended queries may deviate from the business scenario |
collectionIds | Bind dataset ID of the corresponding project | Differentiates knowledge bases for different survey and design projects, prevents cross-project data from interfering with retrieval results |
Three common errors
- Phenomenon: Knowledge base search responses time out, with single-round retrieval taking over 25 seconds. Cause: The
limitparameter is not adjusted for long-text survey reports, and excessive chunk data is recalled in a single round, causing loading delays. - Phenomenon: The exported dataset.csv only includes the index field, with no content field. Cause: Full content extraction is not enabled when importing documents, only document metadata is synced.
- Phenomenon: Tool calls cannot associate knowledge base retrieval results. Cause: The
collectionIdsparameter is not specified, so the retrieval scope covers irrelevant project datasets.
How to verify correct configuration
- A retrieval test is run on a single long-text survey report, and the matching degree of recalled results is confirmed to fall within the preset
similarityrange. - The knowledge base dataset is exported, and the exported file is checked to include both index and content fields, verifying the content extraction configuration is correct.
- The tool call workflow is triggered, and retrieval results are confirmed to match the content of the dataset bound to the specified
collectionIds. - Retrieval response logs are reviewed, and latency is confirmed to be within acceptable limits, with no circular loading or timeout errors.
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.