What data for this category looks like
Petroleum refining technical documents cover operating procedures, production plans, material balance sheets, 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. Data is updated in real time for process parameters, per shift for shift handover records, and daily for material balance and quality reports. Documents include fields such as unit number, distillation range values, interlock logic number, and shift identifier. Common industrial measurement units include ℃, m³, t, and MPa.
What constraints these characteristics impose on knowledge base retrieval and recall
Multiple continuous production units couple via material and energy networks. Raw material properties fluctuate widely. Retrieval and recall must cover cross-unit associated process data, avoiding isolated matching of single-unit documents. Multiple heterogeneous data sources and layered update frequencies require the retrieval system to filter recall results by update time. The system must prioritize matching the latest data generated in real time or on the same day. Documents contain professional metering fields and units. Retrieval must associate fields and units for precise matching, reducing invalid recall results. Interlock ledgers and change management documents have strong logical correlations. Retrieval must support cross-document associated matching, ensuring reviews cover complete change and execution basis.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.82 | Matches the semantic and field precision of professional refining documents, filters low-relevance general documents |
limit | Top 8–12 entries | Covers the volume of associated data in multi-unit coupling scenarios, avoids excessive results exceeding review processing capacity |
usingReRank | Enabled | For multi-source heterogeneous professional refining documents, re-ranking optimizes the sorting priority of associated results |
datasetSearchUsingExtensionQuery | Enabled | Expands the query scope for professional terms such as crude oil evaluation and distillation range control, covering detailed process scenarios |
searchMode | mixedRecall | Balances full-text semantic matching and field precise matching, adapting to the structure of refining documents containing metering fields |
collectionIds | Grouped by business domains such as process, equipment, and laboratory testing | Distinguishes knowledge base collections across different business domains, narrows the retrieval scope, and improves recall precision |
Three common configuration mistakes
- Issue: After upgrading FastGPT v4.17.0, the same query fails to recall previously matched knowledge base documents. Cause: The default
similaritythreshold was adjusted after the version upgrade, or thecollectionIdsconfiguration was not migrated with the original grouping rules. - Issue: After a Docker Compose private deployment, an error prompt appears when clicking the knowledge base module. Cause: The local storage path or vector database port mapping for the knowledge base connection was not correctly configured in the default settings.
- Issue: Knowledge base retrieval response time exceeds the business’s acceptable range, and GPU resources are not fully utilized. Cause: Parallel processing for
usingReRankis not enabled, or thelimitparameter is set too high, resulting in an excessive volume of recalled data.
How to verify configuration correctness
- Submit a query containing refining professional terms. Check whether returned results cover documents from corresponding business domains such as process, equipment, and laboratory testing. Adjust the
similaritythreshold until matching accuracy meets review requirements. - Submit a cross-unit associated query. Verify whether associated process data from multiple units can be recalled. Confirm that the
searchModeandusingReRankconfigurations are active. - View retrieval response time and vector database resource usage. Adjust the
limitparameter to ensure retrieval efficiency meets the business scenario’s requirements. - Test a retrieval request filtered by update time. Confirm whether the latest documents generated in real time, per shift, or daily are recalled first.
The values provided here are common starting points, and should be evaluated against the reader's own samples.