What the data for this category looks like
Data sources for petroleum refining inspection and process record data include DCS/APC control systems, LIMS laboratory systems, tank farm metering systems, crude oil evaluation databases, equipment management systems, and energy management systems. Data update frequencies cover real-time (unit operation parameters), shift-by-shift (shift handover records), and daily (material balance and distillate export quality reports). Daily documents include operating procedures, production plans, interlock ledgers, change management files, and more. Structured documents include fields such as unit number, distillation range values, metering tank number, and interlock trigger time, with units including ℃, m³, ppm, hours, and others.
What constraints do these characteristics impose on the "citation source and traceability" link
Multiple data sources and differentiated update rhythms require the traceability system to match different timestamp verification rules, to avoid associating outdated unit parameters or delayed laboratory data. Diverse document structures require the traceability module to identify the boundary between structured reports and plain text records, to ensure accurate matching of associated fields and units. A production network with coupled multiple units requires traceability to associate logs across systems, rather than being limited to a single document. Otherwise, the root cause of process fluctuations cannot be fully traced. Additionally, the high-risk nature of inspection and audit requires that traceability results have a complete traceable link, and associated information for key links cannot be omitted.
How to set up the configuration
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
quoteMaxToken | 1500–2000 token | Matches the content length of long documents in the petroleum refining industry, such as complete distillate export quality reports |
responseData | Return only metadata + original text snippets | Adapts to traceability requirements for multi-system coupling, avoiding redundant non-associated data |
detail | Enabled | Retains complete document paths, update timestamps, and system source fields to facilitate cross-system association |
| Maximum Context | 8000–12000 characters | Covers the complete content of process logs and shift handover records for multi-unit linkage |
| Single Search Limit | Top 6 entries | Matches the recall volume of multiple data sources, balancing information completeness and retrieval efficiency |
| View Source Original Text | Configured by permissions | Adapts to document viewing permission requirements of different departments, complying with enterprise data security specifications |
Three common mistakes
- Phenomenon: Q&A only returns original text snippets and does not generate corresponding answer content. Reason: Key indicators of structured documents are not extracted as the basis for Q&A pairs, or document field boundaries are not identified during chunking.
- Phenomenon: The returned source field is empty, and specific system documents cannot be associated. Reason: Metadata field mapping for LIMS and DCS systems has not been completed, and system-collected data has not been bound to knowledge base documents.
- Phenomenon: The model test prompts an error, and the citation function works normally after forcibly ignoring it. Reason: The timeout threshold configuration for the
detailparameter is too short; the test request fails to complete full metadata retrieval, but formal requests can be retried.
How to confirm the configuration is complete
- Initiate a query that includes distillation range data, and check whether the returned results include the document number and update time of the corresponding LIMS report.
- Import a shift handover record document, trigger Q&A, and check whether the traceability result is associated with the corresponding unit operation log.
- Adjust the
quoteMaxTokenparameter value, confirm that the citation snippet of long text is not truncated, and complete material balance data is retained. - Check the permission configuration of workflow nodes, confirm that the on/off status of the view source original text switch for different roles meets expectations.
The values provided are common starting points and should be measured against the reader's own samples.