TADI: Agentic LLM for Drilling Intelligence

🔑 Enhanced Key Takeaways

• •TADI addresses the 'data silo' problem in oil and gas by bridging the gap between unstructured Daily Drilling Reports (DDRs) and structured WITSML/production databases, which historically required manual reconciliation.
• •The system utilizes a ReAct (Reasoning and Acting) prompting framework to enable the LLM to dynamically select from the 12 domain tools, reducing hallucinations by forcing the model to cite specific database entries as evidence.
• •The project leverages the Equinor Volve dataset, a widely recognized industry benchmark, to ensure the reproducibility of its findings and to provide a standardized environment for testing agentic workflows in subsurface engineering.

🛠️ Technical Deep Dive

• •Architecture: Hybrid RAG (Retrieval-Augmented Generation) system utilizing DuckDB for SQL-based structured data querying and ChromaDB for vector-based semantic search of narrative reports.
• •Parsing Engine: Custom XML parser designed to normalize heterogeneous schemas across 1,759 files, specifically handling variations in naming conventions and units of measure.
• •Evaluation Metric: The Evidence Grounding Score (EGS) measures the alignment between the LLM's generated response and the retrieved raw data snippets, penalizing 'hallucinated' facts not present in the source documents.
• •Tooling: 12 domain-specific tools include functions for time-series trend analysis, well-bore trajectory verification, and automated cross-referencing of drilling parameters against historical production outcomes.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline