What the data for this category looks like
Data sources for this category include regulatory standard texts, online environmental monitoring data, gas detector readings, operator qualification documents, equipment isolation lists, and reports submitted to supervision platforms. Daily documents cover risk identification lists, work tickets for hot work, confined space operations, lifting operations and other tasks, emission monitoring logs, hazardous waste ledgers, accident rectification reports, and more. Data is divided into three categories: structured forms, long-text regulatory documents, and time-series monitoring data. Fields include work identifiers, monitoring parameter thresholds, qualification numbers, and compliance judgment standards. Units vary by monitoring type: gas concentration uses ppm, emissions use mg/m³. Update frequency differs by scenario: work tickets are generated for each operation, monitoring data is uploaded in real time or on a periodic basis, and regulatory standards are updated on a regular schedule.
What constraints these characteristics impose on knowledge base retrieval and recall
A mix of multiple data types requires retrieval logic to support both structured field matching and long-text semantic recall. This prevents valid content from being missed by a single recall method. Large differences in data update frequencies require distinguishing recall priorities between static regulatory documents and real-time monitoring data. This ensures pre-operation audits obtain the latest monitoring results first. The specificity of fields and units in structured documents requires matching field attributes and units during retrieval. This prevents retrieval failure caused by unit mismatches. A single operation audit needs to associate multiple documents. This requires limiting the retrieval scope and result count to avoid context overload that affects model inference.
The following configurations are based on the knowledge base retrieval logic of FastGPT v4.17.0.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Aligns with documents in this category that have compliance thresholds. Balances retrieval precision and coverage. Avoids introducing irrelevant content if set too low, and misses valid risk items if set too high |
limit | Top 8 entries | Covers the associated documents required for a single operation audit, such as work tickets, monitoring data, and qualification documents. Prevents overly long contexts that exceed model window limits |
searchMode | mixedRecall | Combines keyword matching and semantic recall. Adapts to mixed retrieval needs for structured work tickets and long-text regulations |
usingReRank | Enabled | Rearranges retrieval results by business relevance. Prioritizes returning core documents needed for pre-operation risk identification |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands terms related to the operation. For example, expands "hot work" to corresponding work tickets, gas detection records, and guardian qualifications. Improves retrieval accuracy |
collectionIds | Bind to the dedicated knowledge base collection for safety, environmental protection and special operations management | Limits retrieval scope, eliminates interference from cross-category documents, and ensures only compliant documents from this category are retrieved |
Three common mistakes
- Retrieval returns empty results during conversation, but can recall results in knowledge base tests. Phenomenon: The application returns empty results when called, but matching succeeds in the knowledge base backend test. Cause: The application did not correctly bind the dedicated collection specified by
collectionIds, or thesimilaritythreshold was set too high to filter valid matching content. - Retrieval returns documents unrelated to the query. Phenomenon: Returns hazardous waste ledgers or regulatory texts when the query is unrelated to operation audits. Cause:
datasetSearchUsingExtensionQueryis not enabled, andsearchModeonly relies on semantic recall without combining keyword matching of structured fields. - Mathematical or physical formulas fail to display properly. Phenomenon: Formulas in the knowledge base appear as garbled text or plain text in responses. Cause: The standard format of formulas was not preserved during document import, or text preprocessing rules adapted for formula rendering were not configured.
How to confirm the configuration is correct
- Access the application's knowledge base configuration page, and verify that
collectionIdsis bound to the dedicated knowledge base collection for safety, environmental protection and special operations management. - Submit a typical operation audit query, and check that the
similarityscore of returned results falls within the 0.72–0.80 range. - Import a test document containing formulas, submit a corresponding query, and confirm that formulas are rendered in standard format.
- Submit a query unrelated to operation audits, check that the number of returned results matches the
limitsetting, and no irrelevant cross-category documents appear.
The values provided are common starting points and should be measured against the reader's own samples.