What Data for This Category Looks Like
Infrastructure construction internal office data mainly comes from project construction logs, CAD construction drawings, safety specification manuals, equipment operation and maintenance records, bill of quantities, and similar sources. Data update rhythm follows project progress and policy adjustments, with no fixed cycle. Update frequency is higher during key project phases such as foundation pouring and main structure topping. Document structures include long text paragraphs, structured tables, bitmap-format construction drawings, and on-site photos. Fields mostly include project number, construction area, equipment model, safety level, and similar attributes. Units involve professional engineering units such as meters, cubic meters, and man-hours.
Constraints Imposed on Knowledge Base Retrieval and Recall
The coexistence of long text and structured tables requires the retrieval link to support reasonable chunking and field extraction to avoid semantic loss. Documents with a high proportion of drawings and on-site photos need to rely on image understanding models for indexing, otherwise visual content cannot be retrieved. Data updates have no fixed cycle, so the retrieval system must support incremental synchronization configuration to ensure document timeliness. The total volume of documents for a single project is large, so the number of returned entries per single retrieval must be controlled to avoid context window overflow. For queries dense with professional terms, the semantic matching range must be expanded to improve recall accuracy.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Filter low-relevance results, adapt to retrieval accuracy for long text and structured data in infrastructure projects |
limit | Top 10 entries | Control the volume of returned results to avoid context overflow, adapt to retrieval scenarios with multiple documents per project |
searchMode | Hybrid retrieval | Cover both text keyword and semantic matching, adapt to mixed content types such as construction logs and drawing descriptions |
usingReRank | Enabled, take top 3 entries | Optimize sorting results, improve display priority of highly relevant content, adapt to retrieval needs for mixed sorting of multiple documents |
datasetSearchUsingExtensionQuery | Enabled | Expand the semantic scope of user queries, adapt to query scenarios with many professional terms in infrastructure scenarios |
collectionIds | Fill in the knowledge base collection ID bound to the project | Isolate documents of different projects, avoid interference from cross-project retrieval |
Three Common Misconfigurations
- Issue: No response when uploading multiple documents, and long network failure prompts appear after uploading a single large-volume document. Cause: Infrastructure project documents are mostly PDF or CAD files over 100MB, and chunked upload configuration is not enabled, causing single upload load to exceed system thresholds.
- Issue: No image understanding model option when creating a new knowledge base. Cause: The local deployment did not configure the environment variables for the image indexing model, or the deployment package used did not integrate the dependent components required for this function.
- Issue: Retrieved results do not display the referenced knowledge base source information. Cause: The source association configuration for retrieval results is not enabled, or document metadata fields are not retained during chunking, making it impossible to match source identifiers.
How to Verify Proper Configuration
- Upload a construction plan document that includes a bill of quantities table, initiate a query containing the keyword "quantities", and check whether the returned results include specific field content in the table.
- Upload an on-site photo of foundation pouring, initiate a query containing "on-site pouring", and check whether the returned results are associated with the text index content of the photo.
- Initiate a query containing professional engineering terms, and check whether the sorting of returned results meets expectations, and whether highly relevant content is ranked first.
- View the display area of each retrieval result, and check whether it includes the document source name and upload time.
The values given 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.