What the Data for This Use Case Looks Like
Core data for petroleum refining scenarios comes from daily documents and system data from production operations, process, and scheduling departments. Daily documents include operating procedures, production plans, material balances, distillate export quality reports, interlock ledgers, shift handover records, and change management documents. Data sources cover DCS/APC, LIMS, tank farm metering, crude oil evaluation, equipment management systems, and energy systems. Update frequencies include three categories: real-time, per shift, and daily. Document structures include structured parameter fields and unstructured operation instructions. Field units use industrial standard units such as distillation range (℃), density (kg/m³), and interlock actuation pressure (MPa).
Constraints Imposed on Knowledge Base Retrieval and Recall
Multiple continuous units are coupled via material and energy networks. Retrieval must precisely limit the unit scope to avoid interference from cross-unit irrelevant data. Feedstock properties fluctuate widely. The retrieval system must adapt to process parameters for different crude oils, and support expanding search terms to cover technical terminology. Data comes from multiple heterogeneous sources. Retrieval must support recall logic for both structured ledgers and unstructured documents. Updates occur in real time, per shift, and daily. The retrieval pipeline must support near-real-time data synchronization and incremental recall to avoid returning outdated process information. Document structures are complex. Chunking strategies must adapt to the different length needs of long-text operating procedures and short-entry interlock ledgers.
Configuration Settings
This guide is written for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | High precision requirements for petroleum refining process parameters require filtering low-relevance non-process results |
limit | Top 8–12 entries | Coupled units generate large volumes of process information. Control total context length to avoid exceeding model input limits |
searchMode | mixedRecall | Support simultaneous recall of structured interlock ledgers, distillate export quality reports, and unstructured operating procedures |
usingReRank | Enabled | Wide fluctuations in feedstock properties require reranking to improve the ranking priority of highly relevant process documents |
datasetSearchUsingExtensionQuery | Enabled | Many process technical terms require query optimization to expand search terms, covering scenarios such as crude oil evaluation and unit load adjustment |
collectionIds | Configured per single unit | Process data for multiple units is independent. Precisely limit retrieval scope to avoid interference from cross-unit data |
Three Common Configuration Mistakes
- Embedded images in imported process documents fail to display properly. Symptoms include blank image areas on the knowledge base preview page, with logs showing file not found errors. The cause is only the document itself was uploaded, and image resources were not synchronized to the storage path associated with the knowledge base.
- Knowledge base retrieval takes longer than 10 seconds. Symptoms include delayed interface returns and frontend loading timeouts. The cause is no
collectionIdsconfigured to limit retrieval scope. Full cross-unit data retrieval causes excessive system load. - Retrieval results do not include core process interlock parameters. Symptoms include missing key interlock actuation condition entries in returned results. The cause is the
similaritythreshold is set too high, filtering some highly relevant interlock ledger data with slightly lower similarity scores.
How to Verify Proper Configuration
- Check the document storage path associated with the knowledge base. Upload all embedded image resources, then verify that all images display normally when previewing a single process document.
- Configure
collectionIdsto the dataset ID of a single target unit. Initiate retrieval, then confirm that returned results only include documents such as operating procedures and interlock ledgers for that unit. - Adjust the
similarityandusingReRankparameters. Compare the relevance ranking of retrieval results across different parameter combinations, then confirm that they meet the precision requirements for the process scenario. - Simulate real-time updated process data to trigger retrieval. Confirm that returned results include the latest shift handover records or unit load adjustment plans.
The values provided are common starting points and should be measured against the reader’s own samples.