What this category of data looks like
This category of data comes from production operation, process, scheduling, equipment, and laboratory departments of petroleum refining enterprises. Daily documents include operating procedures, production plans, material balances, distillate export quality reports, interlock ledgers, shift handover records, and change management documents. Data sources include DCS/APC systems, LIMS laboratory management systems, tank farm metering systems, crude oil evaluation databases, equipment management systems, and energy management systems. Update frequencies cover real-time, per-shift, and daily. Documents contain fields such as process parameter values, equipment status, and material indicators, with no unified fixed format, varying by document type.
What constraints these characteristics impose on the "knowledge base retrieval and recall" link
Multi-source heterogeneous data sources and non-uniform document formats require the retrieval link to support cross-system data associated recall and format adaptation. Differentiated update frequencies of real-time, per-shift, and daily require a flexible incremental synchronization and version management mechanism to ensure recalled results are the latest valid data. The high confidence requirements for professional content such as process parameters and interlock logic require improving the precise matching ability of retrieval to avoid irrelevant results interfering with auditing judgments. Multi-unit coupled business scenarios require retrieval to associate cross-unit material and hydrogen energy network data to meet business needs for load coordination and fluctuation handling.
How to set the configurations
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category has a high proportion of professional content. A higher similarity threshold can filter irrelevant non-auditing related documents and ensure the accuracy of recalled results |
limit | Top 6–10 entries | Refining auditing requires covering multi-dimensional associated data. Too many results will exceed the context window, while too few will fail to cover complete business logic |
searchMode | mixedRecall | This category has both semantic matching needs for professional terms and exact matching needs for fixed fields such as interlock number, material indicators. The hybrid mode can balance semantic recall of professional content and exact matching of fixed fields |
usingReRank | Enabled | Semantic similarity of professional documents is easily disturbed by terms. Reranking can further screen the most matching auditing related content and improve result quality |
datasetSearchUsingExtensionQuery | Disabled | The auditing requirements of this category are mostly fixed business scenarios. Extended queries can easily introduce irrelevant cross-unit data and affect auditing accuracy |
collectionIds | Bind the dedicated knowledge base collection for refining inspection and process record auditing | Isolate documents from other business modules to avoid irrelevant content interfering with retrieval results, and ensure the recall scope only covers target auditing data |
Three common errors
- Symptom: Retrieval results include generated content that is not a knowledge base question-and-answer pair, and the displayed answer does not strictly reuse the original knowledge base text. Cause:
datasetSearchUsingExtensionQuerywas not disabled and strict matching mode was not configured, causing the system to generate additional content. - Symptom: The number of recalled results does not match the configured
limitparameter, exceeding or falling short of the expected quantity. Cause:collectionIdswas not correctly configured, causing the retrieval scope to cover non-target knowledge base collections and mixing in irrelevant documents. - Symptom: After long documents are segmented, the semantics of key process parameters are split and cannot be accurately recalled. Cause: The segmentation logic was not adjusted for professional titles, parameter lines, and other delimiters of refining documents, and the default segmentation rules caused semantic breaks.
How to confirm the configuration is correct
- View the retrieval log to confirm that the document source of the recalled results belongs to the knowledge base collection corresponding to the configured
collectionIds. - Enter professional process questions and check that the similarity scores of the recalled results fall within the interval configured for
similarity. - Test entering the original question of the question-and-answer pair stored in the knowledge base and confirm that the returned result is the original answer stored in the knowledge base.
- Simulate real-time data updates and check whether the retrieval results include the latest process parameters or records.
The values provided here are common starting points and should be measured against the reader's own samples.