What data for this category looks like
This category covers the entire production process, with sources including DCS/APC control systems, LIMS laboratory information management systems, tank farm metering systems, crude oil evaluation databases, equipment management systems, and energy management systems. Update frequencies are divided into real-time (process operation parameters), shift-based (shift handover records), and daily (material balance, distillate export quality reports). Daily documents include operating procedures, production plans, material balance records, interlock ledgers, shift handover records, change management documents, and others. Fields cover process parameters, material components, quality indicators, and equipment status, with corresponding units of measurement.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The heterogeneous multi-source data, layered update frequencies, and scenario-based document structure of this category impose multiple constraints on retrieval and recall. Real-time process parameters must be synchronized to the vector database at high frequency to avoid using expired data for defect analysis. Documents updated on a shift-based or daily basis require scheduled synchronization tasks to ensure retrieval results match the latest production requirements. Document formats vary widely across different sources, so chunking rules must be set separately for time-series process data and structured quality reports. Additionally, the production logic of coupled multiple units requires that retrieval results be associated with relevant documents from upstream and downstream units to improve the completeness of traceability.
How to set the configuration
This page applies to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Filter low-match non-process results to meet the precise retrieval needs of quality traceability |
limit | Top 8–12 entries | Balance recall coverage and context length limits to meet the needs of multi-unit associated retrieval |
usingReRank | Enabled | Original semantic matching may have deviations under heterogeneous data; reranking improves the accuracy of process indicator matching |
datasetSearchUsingExtensionQuery | Enabled | Can supplement extended query terms such as crude oil properties and unit load to adapt to production scenarios with large raw material fluctuations |
searchMode | embedding | Match the semantic association between process logic and quality indicators to adapt to scenario-based traceability needs |
collectionIds | Bound by production unit classification | Limit the retrieval scope to the knowledge base collection of associated units to reduce interference from irrelevant results |
Three common mistakes
- Phenomenon: Semantic retrieval returns results with high matching scores, but they are actually irrelevant to the current quality problem. Cause: No reasonable
similaritythreshold is set, andusingReRankis not enabled, resulting in high-score non-process results being recalled first. - Phenomenon: When referencing a knowledge base ID or system variable, retrieval results do not change or an error is triggered. Cause: The
collectionIdsparameter is not configured correctly, the variable is not bound to the specified knowledge base collection, or the variable format does not meet system requirements. - Phenomenon: After importing an Excel spreadsheet dataset, retrieval cannot match the field correspondence in the table. Cause: Specialized chunking and indexing rules are not configured for structured spreadsheet data, resulting in lost field association relationships.
How to confirm the configuration is correct
- Submit a query containing specific process parameters and quality indicators, and check whether the matching scores of the returned results fall within the preset
similarityrange. - After binding the
collectionIdsof a specified unit, verify that the retrieval results only include documents from that unit and its associated units. - After enabling
datasetSearchUsingExtensionQuery, submit a query that does not explicitly mention crude oil properties, and check whether the returned results supplement relevant associated content. - After importing an Excel structured dataset, retrieve a specific field value in the table and verify that the corresponding entry can be recalled.
The values provided are common starting points and should be measured against the reader's own samples.