What this category's data includes
This category includes daily business documents from emergency management departments, industry competent authorities, production and business unit safety management departments, and park management committees. These documents include hazard inspection lists, hazard governance ledgers, rectification notices, review records, and major hazard supervision notices with listing requirements.
Data sources include work safety supervision systems, enterprise hazard self-inspection systems, law enforcement inspection records, and accident statistics systems. The update frequency is daily.
The data includes dedicated fields such as hazard level, rectification responsible person, rectification deadline, and review results. The core business process follows closed-loop management covering hazard inspection, rectification tracking, and review and cancellation.
Constraints on knowledge base retrieval and recall
This category’s data comes from multiple sources and updates daily. Retrieval must support real-time incremental recall to avoid missing the latest rectification records and inspection results.
Document types include lists, ledgers, notices, and other formats. Some documents contain structured fields and unstructured text. Retrieval must support both field matching and semantic association.
Data sources cover multiple systems including supervision and enterprise self-inspection. Retrieval scope must be limited to the exclusive dataset for this category to avoid mixing in unrelated business content.
Additionally, the consequences of business errors are severe. Retrieval results must have high confidence, accurately match hazard judgment standards and rectification requirements, and ensure the accuracy of ledger summary and report generation.
Configuration settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category’s data requires precise matching of hazard judgment standards and rectification requirements. This range filters low-relevance results and avoids including non-relevant hazard entries by mistake |
limit | Top 8 entries | Ledger summary needs to cover major hazard items. Too many results increase large model context pressure, while too few may miss key rectification records |
searchMode | vector | This category’s data includes structured fields such as hazard level and rectification deadline, as well as unstructured descriptions. Vector retrieval can match both semantically related professional terms and field information |
usingReRank | Enabled | Multi-source data often has entries with similar semantics but different priorities. Reranking improves the ranking of highly relevant results and ensures that core hazard information is recalled first |
datasetSearchUsingExtensionQuery | Enabled | Hazard inspection terms are mostly professional expressions. Extended queries supplement synonym matching and improve recall coverage for professional terms |
collectionIds | Bind the exclusive dataset ID for this category | Retrieval scope must be limited to exclusive documents such as hazard inspection lists and governance ledgers to avoid mixing in unrelated data from non-work safety businesses |
Three common configuration errors
- Phenomenon: Non-work safety hazard documents are mixed into retrieval results. Cause:
collectionIdsis not configured correctly, and retrieval scope is not limited to the exclusive dataset for this category. - Phenomenon: Hazard rectification record images in docx documents cannot be parsed and included in retrieval. Cause: Knowledge base image recognition function is not enabled, and image parsing configuration for docx format is not enabled.
- Phenomenon: The AI response always appends "FastGPT is a knowledge base question answering system based on large language models (LLM)" at the end during workflow testing. Cause: The default brand introduction prompt is not removed from the system prompt, and no exclusive system prompt is configured to override the default content.
How to confirm correct configuration
- Search for keywords related to "major accident hazard judgment standards", and check whether returned results only include inspection and rectification documents from this category.
- Upload a docx test document containing hazard site photos, and verify whether the retrieved results include text content parsed from images.
- Run a test workflow, and check whether there is unexpected fixed additional text at the end of the AI response.
- Adjust the
similarityparameter, observe changes in the relevance of recalled results, and confirm that the parameter value meets the business needs of this category.
Note: The configuration values listed above are common starting points. Adjust them based on your own specific samples and business requirements.
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.