RAG Skill

RAG (Retrieval-Augmented Generation) Skill

This skill provides retrieval tools for accessing simulator documentation and example files through vector search.

Overview

The RAG skill enables agents to:

• Retrieve official simulator manual documentation for keyword definitions and syntax
• Retrieve example DATA files and case studies for concrete usage patterns
• Answer keyword format questions with authoritative sources
• Provide working examples to illustrate keyword usage

Tools

simulator_manual

Retrieves information from the simulator manual and official documentation using semantic search.

Usage:

simulator_manual(query: str) -> str

Parameters:

• query: Natural language query about keywords, syntax, or documentation (e.g., "COMPDAT keyword format", "WELSPECS syntax and fields")

Example:

simulator_manual("What is the COMPDAT keyword format?")

Returns: Retrieved documentation snippets from the simulator manual with source citations.

When to use:

• First step in keyword Q&A flow (TOOL_DECISION_TREE.md Section 2.5)
• In scenario test chain after parse_simulation_input_file (Section 2.4) to get keyword context
• When user asks about keyword definitions, syntax, or parameter tables

simulator_examples

Retrieves example DATA files and case studies using semantic search.

Usage:

simulator_examples(query: str) -> str

Parameters:

• query: Natural language query about keyword examples or usage patterns (e.g., "COMPDAT keyword format", "WCONINJE injection rate examples")

Example:

simulator_examples("COMPDAT keyword format")

Returns: Retrieved example DATA file snippets showing concrete keyword usage.

When to use:

• After simulator_manual when manual lacks format details or examples (Section 2.5)
• In scenario test chain after simulator_manual to get example context for modifications (Section 2.4)
• When user needs working examples to understand keyword syntax

Workflow Integration

This skill integrates with the Simulator Agent's decision tree (TOOL_DECISION_TREE.md):

• Keyword Q&A Flow (Section 2.5):

   simulator_manual → (simulator_examples) → answer

• Scenario Test Chain (Section 2.4):

   parse_simulation_input_file → simulator_manual (inferred keyword) → simulator_examples (same keyword) → modify_simulation_input_file → run_and_heal

• Keyword Chain (Section 3.4):

   simulator_manual → simulator_examples → synthesize format + example → final answer

Implementation Details

Tools are implemented as LangChain retriever tools with:

• Vector store: Milvus containing embedded simulator manual and examples
• Embeddings: NVIDIA embedding API for semantic search
• Top-k retrieval: Configurable number of relevant chunks returned (default: 10, reranked to 5)
• Source citation: Metadata includes source file paths and section references

extract_keyword (internal helper)

scripts/extract_keyword.py provides RAG + LLM keyword extraction (e.g. "plot field oil" → FOPT). Used by plot_skill validators and other skills that need to infer keywords from natural language.

from simulator_agent.skills.rag_skill.scripts.extract_keyword import extract_keyword

kw = extract_keyword("plot field cumulative oil production", intent="summary_metric")  # -> "FOPT"

Best Practices

• Always cite sources: Include manual section names and example file paths in responses
• Use simulator_manual first: It's the authoritative source for syntax and fields
• Use simulator_examples for illustration: Examples show concrete usage but should not override manual definitions
• Handle conflicts: If examples conflict with manual, prefer manual and explain the conflict

References

• Tool Decision Tree - Routing logic
• OPM Flow Manual - Manual structure and indexing

Configuration

RAG tools use Milvus collections created by ./scripts/setup.sh --full:

| Tool name | Milvus collection | Ingested by |

|---------------------|---------------------------|--------------------------------|

| simulator_manual | docs | ingest_papers.sh |

| simulator_examples| simulator_input_examples| ingest_opm_examples.py |

Environment variables:

• MILVUS_URI: Milvus endpoint (default: http://localhost:19530; Docker: http://standalone:19530)
• NVIDIA_API_KEY: Required for embeddings, reranker, and LLM