This document was written for FastGPT v4.17.0.
What this category of data looks like
Petroleum refining scenario data originates from multiple systems. These include unit operation parameters collected in real time by DCS/APC, distillate quality reports generated by LIMS, material data measured by tank farms, crude oil evaluation reports, equipment management ledgers, and energy consumption data from energy systems. Daily documents include operating procedures, production plans, material balance sheets, interlock ledgers, shift handover records, and change management documents. Data update cycles cover real-time, shift-based, and daily frequencies. Fields include crude oil density, distillation range, product sulfur content, unit load, and corresponding engineering units.
What constraints these characteristics impose on model access and configuration
Real-time and batch updates of multi-source heterogeneous data require model access to support streaming data pulling and scheduled incremental synchronization. This avoids full data pulling that occupies excessive system resources. Diverse document types and fields require configuration of parsing rules adapted to different data structures. This ensures accurate identification of fields such as sulfur content in quality reports and unit interlock logs. Continuously coupled unit processes require model context to cover multi-link associated data. It is also necessary to limit the number of tokens processed per round to prevent request timeouts. In addition, defect analysis in refining scenarios requires association with historical interlock and material balance data. Configuration items must support associated recall of multi-source data.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Covers multi-unit coupled process data and historical analysis documents, avoids truncation of critical context |
maxTokens | 4000 token | Controls single-round response length, adapts to the detailed output requirements of refining analysis reports, prevents timeouts |
quoteMaxToken | 1500 characters | Limits the single-segment length of recalled documents, matches the field length of single records in refining quality reports |
| Workflow/API tool selection | auto | Allows the model to independently select tool calls, adapts to the associated query requirements of multi-source data |
temperature | 0.3–0.5 | Reduces random generation probability, ensures consistency and traceability of defect analysis results |
AIPROXY_API_TOKEN | Key matching the deployed model | Completes identity verification for model access, prevents unauthorized requests from being blocked |
AIPROXY_API_ENDPOINT | Fill in according to the deployment environment | Adapts to Ollama or other locally deployed model access addresses, ensures model call connectivity |
Three common configuration mistakes
- When
toolChoiceis set tonone, the model cannot trigger tool calls. This manifests as an inability to return real-time unit data or historical interlock records. The cause is that tool call permissions are forcibly disabled, which does not adapt to the requirement of associating multi-system data in refining scenarios. - When
maxContextexceeds 16000 token, model calls return a 504 timeout error. The cause is insufficient resources of the locally deployed Ollama, which cannot handle processing requests with overly long context. - When
AIPROXY_API_TOKENis not configured, model access requests return a 401 unauthorized error. The cause is that the identity verification token for model access is not correctly filled, resulting in request blocking.
How to confirm successful configuration
- Initiate a test request that includes queries for multiple unit parameters. Check whether the model can recall real-time data from the corresponding DCS system to verify that the tool call function works properly.
- Submit a quality report analysis request that includes long text. Check whether the response is within the preset
maxTokensrange, with no truncation or timeout. - View the model access logs. Confirm that requests correctly carry the
AIPROXY_API_TOKENand return a 200 status code to verify connectivity. - Adjust the
temperatureparameter and initiate repeated tests. Check the consistency of analysis results to confirm that the parameter takes effect.
The values provided here are common starting points. They should be evaluated against the reader's own samples.