SimVault — graph RAG + semantic search on Simscape/Simulink .slx files (probably the first of its kind)

[nightfury1802]

Hi everyone,

I've been using the MATLAB MCP server heavily for agentic Simscape/Simulink modeling and ran into a recurring problem: every agent session rebuilt the same models from scratch and made the same wiring mistakes. Port names like omega_rpm, w_shaft, SpeedPS are all the same signal — but no agent could know that without reading the .slx XML itself. Domain mismatches (Simscape physical port → Simulink signal line) caused silent errors that only appeared at simulation time.

So I built SimVault — a companion MCP server that indexes your .slx library and gives agents the port-level ground truth they need before calling model_edit.

What it does

SimVault indexes a directory of .slx files into a combined vector + graph store with two layers:

1. Model Graph — port-level structure and compatibility of every indexed subsystem (28 models, 129 ports, 1541 edges in my library)
2. KB Knowledge Graph — session logs, pitfalls, validated results, and design decisions extracted from past Claude Code sessions — linked back to the model nodes via cross-edges

The agent workflow becomes:

simvault_search("PMSM thermal model, high-fidelity, continuous")
    → MotorThermal11Node + PMSM_FEM returned as compatible partners

simvault_get_assembly_context(["PMSM_FEM", "MotorThermal11Node"])
    → canonical port names, solver contracts, pre-validated wire pairs

simulink_mcp.model_edit(...)
    → builds the model using SimVault's context — no guessing

The 6 MCP tools

Tool	What it does
simvault_search	Semantic search with hard fidelity / analysis-type / solver filters
simvault_validate_wire	Returns PASS / WARN / BLOCK on a proposed port connection
simvault_get_assembly_context	Full canonical port table + pre-validated wire pairs for a set of subsystems
simvault_smoke_test	Load check + optional 0.1 s simulation at rated operating point
simvault_kb_query	Unified KB search: semantic + BM25 + graph traversal, fused with RRF
simvault_model_context_with_history	Assembly context + KB session history in one call

Wire validation rules (applied before any model_edit)

Four rules, checked in order:

Rule	Outcome
Domain mismatch (Simscape physical ↔ Simulink signal)	BLOCK
Direction conflict (output → output)	BLOCK
Unit mismatch with known conversion (e.g. rpm → rad/s)	WARN + inserts a Gain block
Solver contract mismatch (continuous ↔ discrete)	WARN + inserts a Rate Transition

Port canonicalization

The canonicalizer maps 20+ raw naming variants to a fixed vocabulary so the graph can reason about compatibility:

• omega, w_shaft, SpeedPS, wr, speed_ref → omega_shaft_rads
• torque, TPS, trq, tau, torque_ref → torque_shaft_Nm
• temp, AirGap, winding_K → temperature_K
• SL2PS_* and PS2SL_* block names detected automatically

KB memory with confidence decay

Each Claude Code session is automatically extracted (via a Stop hook) and added to the KB. Facts are tagged by tier:

Tier	Decay	Examples
working	7 days	Raw session observations
episodic	90 days	graphify cache chunks, session exports
semantic	Permanent	Validated results, design decisions
procedural	Permanent	Model specs, MATLAB/Python source

Superseded or expired facts score 0 in retrieval — agents don't get confused by stale context.

Installation

pip install -e .
matlab -batch "run('examples/tag_models_for_simvault.m')"
simvault index examples/pmsm_drive/

Register alongside the MATLAB MCP server in claude_desktop_config.json:

"simvault": {
  "command": "python",
  "args": ["-m", "simvault.mcp_server"],
  "cwd": "/path/to/SimVault",
  "env": { "SIMVAULT_ROOT": "/path/to/SimVault" }
}

The Stop hook auto-updates the KB at the end of every session — the library compounds over time.

Repo

https://github.com/nightfury1802/SimVault

Python ≥ 3.11, MATLAB R2024a+ for extraction (Python pipeline runs without MATLAB if JSONs already exist).

Happy to answer questions or discuss integration patterns — especially interested in hearing from anyone building multi-agent Simulink workflows where model reuse across sessions is important.

[sylvain-mw]

@derrick-mw Drawing your attention to this.