What This Category of Data Looks Like
Data comes from EHS departments and maintenance teams of chemical, metallurgical, machinery and other manufacturing enterprises. It covers risk identification records, work permits such as hot work, confined space work, lifting work, online environmental monitoring data, hazardous waste ledgers, emission monitoring reports, accident and hidden danger rectification reports. Data sources include industry regulatory standards, online environmental monitoring systems, gas detectors, personnel qualification files, equipment isolation lists and supervision platforms. Updates are triggered by work process events, updated in real time per monitoring requirements, or synced via scheduled inspections. Document fields include work type, risk level, monitoring value, rectification closure status and more. There are no fixed unified units; fields and units adjust dynamically based on document type.
Constraints Imposed on Knowledge Base Retrieval and Recall
This category spans a wide range of data types. It includes both structured real-time monitoring values and unstructured work permit text. Retrieval must balance exact matching and semantic association to avoid missing key compliance information with a single matching method. Data update frequencies vary. Real-time monitoring data requires high-frequency recall. Work permits generated for single jobs need immediate sync triggered by workflow. Data spans departments and processes. Retrieval results must cover multiple dimensions including equipment isolation, personnel qualifications and compliance requirements. Single-dimensional recall may lead to missing key compliance items in work permits. Some data involves safety and environmental veto points. The confidence of recall results must meet business requirements. Irrelevant or incorrect content that affects decision-making must not appear.
How to Set Configurations
| Configuration Item | Recommended Value | Basis for This Value |
|---|---|---|
similarity | 0.75–0.85 | This category involves safety and compliance content, so low-match results must be filtered to avoid irrelevant information interfering with decision-making |
limit | top 6 entries | Multiple types of information must be covered, including work permits, monitoring data and personnel qualifications. 6 entries balances retrieval breadth and context length within 2000 tokens |
searchMode | mixedRecall | Balances semantic matching for work permit text and exact recall of regulatory keywords, adapting to multiple data types |
usingReRank | enabled | This category has high data relevance requirements. Reranking optimizes result sorting to ensure high-priority compliance items are displayed first |
datasetSearchUsingExtensionQuery | enabled | Can automatically expand compliance-related keywords based on user queries, covering implicit requirements of work permits |
collectionIds | bind exclusive collections by business domain | Isolates safety, environmental protection and production quality data to avoid compliance risks from cross-category retrieval |
Three Common Misconfigurations
- Phenomenon: Retrieval results only return the top 1 entry, and it is not possible to select the top 2 entries for processing by the large model. Cause: The value range of the
limitparameter is not adjusted, and the context splicing logic when calling the large model is not configured. - Phenomenon: Index update cannot be triggered after adding data to a collection, and an error prompts that a training order is missing. Cause: Confused between incremental synchronization of collection data addition and full/incremental index construction process of training orders, and did not trigger formal index generation via a training order.
- Phenomenon: Retrieval results have low matching accuracy and fail to hit implicit compliance requirements in work permits. Cause:
datasetSearchUsingExtensionQueryis not enabled, and user queries are not pre-optimized, resulting in failure to cover business-related keywords.
How to Confirm Configuration Is Successfully Applied
- Enter the knowledge base configuration page of FastGPT v4.17.0, and check whether the values of parameters such as
similarity,limit,searchModematch the configuration table. - Simulate submitting a query related to hot work, and check whether the number of returned results conforms to the
limitsetting, and whether the sorting is optimized via reranking. - Upload a new work permit document, and check whether a training order is generated and index update is completed.
- Test a query containing compliance keywords, and confirm that the recall results include corresponding regulations and work requirement content.
The values provided are common starting points and should be evaluated against the reader's own samples.