The RAG system that grounds SimPilot's AI in proven simulation approaches.
SimPilot's knowledge base is a Retrieval-Augmented Generation (RAG) system that ensures the AI agent generates simulation configurations grounded in proven, tested approaches rather than hallucinated settings. It includes specialized domain knowledge, a cross-session memory system, a skills system, and a web search fallback.
How RAG works
When the agent needs domain knowledge, it follows a three-step retrieval process:
Query formulation
The agent constructs a search query based on the current task -- solver type, physics involved, geometry description, or specific error message.
Semantic search
The query is embedded and searched against the knowledge base using semantic similarity. The database returns the most relevant documents, scored by relevance. An LLM-based reranker then normalizes scores across knowledge sources to produce the final ranking.
Context injection
The retrieved documents are compressed and injected into the agent's context as reference material. The LLM uses these documents to inform its decisions, citing specific tutorial configurations and solver settings.
Intent-driven retrieval
Before any search is performed, an intent classifier determines how much retrieval is needed. Requests that are simple reruns or reports may skip retrieval entirely, while new simulation setups trigger full-depth knowledge base queries. The retrieval planner routes queries to the appropriate domain-specific knowledge sources based on the task at hand -- geometry generation, mesh generation, simulation setup, or post-processing.
Knowledge coverage
SimPilot's knowledge base covers solver tutorials, configuration templates, execution scripts, command documentation, mesh generation patterns, CAD geometry examples, and post-processing recipes. Each knowledge source serves a specific role in the workflow pipeline -- from identifying the right solver and directory layout for a new simulation, to providing full configuration file examples, correct command sequences, and visualization techniques.
When knowledge is used
The knowledge base is queried at specific points in the simulation workflow:
Skills system
Beyond the structured knowledge base, SimPilot includes a skills system for domain-specific expertise. Skills are specialized knowledge modules that the agent loads when it encounters tasks requiring domain expertise -- mesh refinement, turbulence modeling, combustion, structural analysis, and more.
How skills work
On-demand loading -- Skills are not loaded into context by default. The agent identifies when specialized knowledge is needed and loads the relevant skill document.
Tagged metadata -- Each skill is tagged with solver type, task category, and keywords, allowing the agent to quickly identify the most relevant skill for the current task.
Composable -- Multiple skills can be loaded simultaneously for tasks that span multiple domains (e.g., mesh generation + turbulence modeling).
Extensible knowledge
The skills system is designed to be extensible. New skill documents can be added to cover emerging solver configurations, specialized physics models, or organization-specific best practices.
Cross-session memory
SimPilot maintains a persistent memory of user preferences and learned patterns across chat sessions. After each substantive session (one that used simulation tools), the system extracts reusable facts and stores them as vector embeddings.
Memory types:
Type
Examples
Solver preference
"User prefers kOmegaSST for external aerodynamics"
Parameter insight
"Relaxation factor of 0.3 works well for this geometry type"
Error pattern
"Recurring SIGFPE resolved by reducing deltaT to 1e-5"
Project context
"Working on automotive underbody simulations at Re=2M"
At the start of each new session, relevant memories are retrieved by semantic similarity and injected into the agent's context. This allows the AI to personalize responses without the user repeating preferences.
Web search fallback
When the local knowledge base is insufficient, the agent has access to the broader internet through web search:
Returns relevant page content with source citations
Searches solver forums, community sites, and official documentation for edge-case solutions
Web search is deliberately positioned as a fallback rather than a primary source. The local knowledge base contains curated, validated configurations, while web results may include outdated or incorrect advice. The agent applies additional scrutiny to web-sourced fixes before applying them.