What the data for this category looks like
Data for petroleum refining scenarios comes from DCS/APC, LIMS, tank farm metrology systems, crude oil evaluation systems, equipment management systems, and energy systems. Daily documents include operating procedures, production plans, material balances, distillate outlet quality reports, interlock logs, shift handover records, and change management documents.
Data update cycles are divided into real-time, per-shift, and daily. Distillate outlet quality reports and shift handover records are mostly updated per shift. Crude oil evaluation and change management documents are updated on demand.
Documents contain specialized fields such as process parameters, material components, and equipment status. Some documents include unit labels, such as density in kilograms per cubic meter and flow rate in tons per hour.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Data from multiple systems is scattered across different business collections. Supplier and process quality audits need to cover full-process data, so retrieval must support cross-collection recall.
Real-time, per-shift, and daily update cycles require retrieval configurations to adapt to incremental synchronization logic, preventing recall of outdated production data.
Documents contain many specialized process fields and units. Vector training must match the precision of business terminology to avoid semantic deviation.
The business characteristic of multiple units coupled via logistics and energy networks requires retrieval results to associate related data from upstream and downstream units, avoiding one-sided judgments of audit basis.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Petroleum refining process parameters have high precision requirements. Low-match irrelevant documents must be filtered to ensure the accuracy of audit basis |
limit | Top 6 entries | Multi-unit coupled business needs to cover upstream and downstream related data, while avoiding context overload that affects large model processing |
searchMode | mixedRecall | Balances semantic matching of professional terminology and precise keyword recall, adapting to the structured and semi-structured characteristics of process documents |
usingReRank | Enabled | Multi-source data is prone to redundancy. Reranking can improve the ranking priority of high-match results and optimize the presentation order of audit basis |
datasetSearchUsingExtensionQuery | Enabled | Supplement business-related derived queries to cover scenarios such as unit load adjustment and quality deviation handling after crude oil property fluctuations |
collectionIds | Bind specified collection IDs including LIMS, DCS, crude oil evaluation, etc. | Precisely limit the retrieval scope to core business collections, avoiding irrelevant system data interfering with the audit process |
Three common mistakes
- Phenomenon: The number of retrieval results does not match the preset
limitparameter, or the specified number of results is not passed to the large model. Cause: The value of thelimitparameter is not correctly passed when calling the interface, or the logic for automatically passing results to the large model is not configured. - Phenomenon: Real-time updated DCS and tank farm metrology data is not indexed, and training tasks remain in a pending state for a long time. Cause: Adding data to a collection is mistakenly equated with creating a training order, and incremental synchronization is not triggered via the training order interface, resulting in real-time data not being vectorized.
- Phenomenon: Retrieval results do not return associated image content, or the business documents corresponding to the images cannot be loaded. Cause: The image vectorization configuration is not enabled, and the recall logic for associated images is not specified in the retrieval rules, only text content is retrieved.
How to confirm the configuration is properly set
- Enter the knowledge base configuration panel of FastGPT v4.17.0, and check whether the values of parameters such as
similarityandlimitmatch the preset configuration. - Upload a latest distillate outlet quality report to trigger a training order, and confirm that the index generation status is completed.
- Input "Unit load adjustment plan under a certain crude oil property" as the query, and the retrieval results should include documents from corresponding collections such as crude oil evaluation and production plans.
- Test calling the retrieval interface, and confirm that the number of returned results matches the value of the
limitparameter.
The values given are common starting points and should be measured against the reader's own samples.