What the data for this category looks like
Petroleum refining scenario business data covers production operation, process, equipment and other departments of oil refining enterprises. Daily documents include operating procedures, production plans, material balances, distillation outlet quality reports, interlocking ledgers, shift handover records, change management documents. Data sources include DCS/APC systems, LIMS systems, tank farm metering systems, crude oil evaluation reports, equipment management systems, energy management systems. Data update rhythms are divided into real-time (DCS time-series data), per-shift (shift handover records), daily (material balance and quality reports). Document fields include standardized unit fields such as unit number, operating pressure (MPa), temperature (℃), equipment serial number, operating duration (hours), and some long documents have multi-chapter nested structures.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Multi-source and heterogeneous data sources require retrieval to be compatible with different formats such as time-series data, static ledgers, and report documents, and targeted chunking and recall rules need to be configured. Different update frequencies require support for a synchronization strategy combining incremental update and full update to avoid data lag or redundancy. Documents in the refining scenario have associated logic of multi-unit coupling. During retrieval, relevant entries of the same unit and same process section must be recalled first, and the context length of recall results must be controlled to avoid exceeding the LLM processing limit. In addition, the consequences of abnormalities in this scenario are severe, so the accuracy of recall results must be ensured to avoid irrelevant entries interfering with decision-making.
How to configure the settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Match fault descriptions for refining equipment inspection with standardized knowledge base entries, filter irrelevant non-refining equipment data |
limit | Top 8 entries | Adapt to the multi-unit coupled business scenario, control the total character count of recall results, and comply with the context limit requirements of FastGPT v4.17.0 |
searchMode | fullTextRecall | Adapt to the retrieval needs of long documents such as production plans and material balances, covering cross-chapter associated process content |
usingReRank | Enabled | Perform secondary sorting on multi-source heterogeneous recall results to improve the probability of accurately matching refining process questions |
datasetSearchUsingExtensionQuery | Enabled | Perform semantic optimization on colloquial inspection questions to match standardized operating procedures and maintenance records |
collectionIds | Configured according to the exclusive knowledge base ID for refining | Isolate knowledge base data of other categories, avoid recalling irrelevant content, and achieve precise knowledge base range limitation |
Three common configuration mistakes
- Misconfiguring
collectionIdsby introducing knowledge base IDs of other categories, resulting in recall of non-refining equipment maintenance entries and search results deviating from the refining scenario. This occurs because the parameter is not configured according to the refining-exclusive collection, leading to mixed cross-category data. - Failing to enable
datasetSearchUsingExtensionQuery. When using colloquial inspection questions such as "How to repair a leaking hydrogenation unit heat exchanger", standardized maintenance procedures cannot be matched, resulting in empty search results. This occurs because semantic expansion optimization is not performed for colloquial questions. - Failing to adjust the
similaritythreshold. When the threshold is too low, a large number of redundant shift handover records are recalled; when the threshold is too high, matching fault maintenance entries cannot be recalled, resulting in abnormal result counts. This occurs because the threshold is not calibrated based on the document characteristics of the refining scenario.
How to verify the configuration is correct
- Initiate questions related to equipment inspection, check that recall results only include operating procedures and maintenance records of refining equipment, with no content from other categories.
- View the FastGPT retrieval log to confirm that configuration parameters such as
similarityandusingReRankare correctly called, and the parameter values comply with preset rules. - Upload a new refining equipment change management document, trigger incremental update, then retrieve the keywords of this document to confirm that the corresponding entry can be normally recalled.
- Test colloquial maintenance questions, confirm that the corresponding standardized document entries can be matched, and verify that the semantic expansion optimization takes effect.
The values provided are common starting points and should be measured against the reader's own samples.