What this category of data looks like
The data for this category mainly comes from the engineering quality and safety supervision system, on-site construction real-name platform, supervision logs, and test reports. The update frequency is weekly. Daily processed documents include supervision plans, supervision inspection records, hidden danger rectification notices, quality problem handling forms, and completion acceptance supervision reports. Each document typically contains fields such as project unique identifier, inspection period, hidden danger description, and rectification requirements. There is no fixed format that is universal across categories, so documents must be split and archived based on business scenarios.
Constraints on knowledge base retrieval and recall
Weekly updated data requires the knowledge base to regularly synchronize incrementally archived content. This prevents expired hidden danger rectification records from being recalled first, which would reduce the accuracy of ledger summary. Documents with multiple types and no fixed formats require the retrieval link to adapt to different text structures, and cannot rely on a single field matching rule. Dispersed multiple data sources require configuring cross-dataset recall logic to ensure coverage of all business-related documents. Because the consequences of business errors are severe, recall results must avoid omitting key hidden danger information. Retrieval matching weights and sorting rules need to be adjusted accordingly.
How to configure the system
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Ensure that the recalled content meets the matching degree requirements for core business fields such as hidden danger description and rectification requirements, and avoid low-relevance content from entering the context |
limit | Top 8 entries | Balance the number of reference documents required for ledger summary and the context length limit, and cover common weekly inspection records and rectification orders for a single project |
usingReRank | Enabled | Re-rank the initially recalled results to prioritize recalling high-value documents containing hidden danger levels and rectification deadlines, adapting to core business retrieval needs |
datasetSearchUsingExtensionQuery | Enabled | Automatically expand retrieval terms, supplement synonymous expressions for fields such as project number and inspection period, and adapt to document naming differences across dispersed data sources |
collectionIds | Configured by project group | Targetedly retrieve exclusive datasets such as supervision plans and inspection records for corresponding projects, avoiding recall of irrelevant documents across projects |
Three common misconfigurations
- Phenomenon: The retrieval debug page only returns 1 result, which cannot meet the need to splice multiple reference documents for ledger summary. Cause: The
limitparameter was incorrectly set to 1, and the parameter was not adjusted to match the business requirement for multiple associated documents. - Phenomenon: Retrieval results include engineering documents from non-current projects, leading to mixed ledger data. Cause: The
collectionIdstargeted retrieval configuration for the corresponding project dataset was not enabled, and full dataset retrieval was used by mistake. - Phenomenon: When retrieving keywords for test reports, image attachment content cannot be recalled, or returned results do not include image-related information. Cause: The knowledge base's image vectorization and multimodal retrieval configuration was not enabled, and only text content was vectorized.
How to confirm the configuration is correct
- Enter the FastGPT knowledge base retrieval debug page, enter the hidden danger description keyword for a single project, and verify that the project identifier of the returned results matches the target project.
- Adjust the value range of
similarity, compare the number of recalled results and their relevance across different values, and confirm that they meet business matching requirements. - Test retrieval of keywords containing project numbers, and verify that
datasetSearchUsingExtensionQueryautomatically expands synonymous retrieval terms to cover documents with different naming formats. - Trigger the re-ranking function, and check whether the sorted recalled results prioritize displaying documents with core fields such as hidden danger levels and rectification deadlines.
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.