Multi-agent LLM workflow that applies the TRIZ methodology to complex engineering problems. Implements a supervisor-routed LangGraph of specialized agents (Project Manager, TRIZ Specialist, Mechanical/Electrical/Control/Safety Engineers, Operations & Documentation), plus TRIZ-specific tools (features list, contradiction matrix, inventive principles) and an optional TRIZ RAG.

The accompanying paper see: TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation

• Agent workflow (LangGraph): Supervisor (Project Manager) routes work among domain agents, tool calls are separate tool nodes.

• LLM providers via LangChain: OpenAI (GPT-4o, o*), Google Gemini (1.5/2.5), DeepSeek, xAI (Grok). Easy to add more.

• TRIZ tools:

  • 39 features list
  • Contradiction matrix → suggested inventive principles
  • Inventive principles lookup
  • Optional TRIZ RAG over PDFs (Chroma + OpenAI Embeddings)

• Experiments: One-shot or batch runs, JSON/Markdown artifacts.

• Evaluation: Multi-judge scoring (clarity, coherence, coverage, novelty, feasibility, TRIZ-adherence, expert-alignment).

• Notebook: quickstart.ipynb for end-to-end runs + graph visualization.

Use our curated requirements.txt:

pip install -r requirements.txt

pip install -e .

Notes:

• On Windows, faiss-cpu is skipped; Chroma will still work.

• For local graph rendering, also install:

  pip install graphviz pygraphviz  # or 'pydot'
  # and install Graphviz system binaries via apt/brew/choco

Create a .env in repo root (copy from .env.example):

# Providers (fill only what you need)
OPENAI_API_KEY=sk-...
GOOGLE_API_KEY=AIza...
XAI_API_KEY=xai-...
DEEPSEEK_API_KEY=sk-deepseek-...

# LangChain Hub (for pulling prompts used by prompt_templates.py)
LANGCHAIN_API_KEY=ls_...

# Tavily (web search tool)
TAVILY_API_KEY=tvly-...

The notebook and modules use python-dotenv to load this automatically.

Open notebooks/quickstart.ipynb and run the cells:

• Installs (optional)
• Loads .env
• Lets you pick the agent model and evaluator models
• Visualizes the LangGraph (offline Graphviz first; Mermaid fallback)
• Runs a full TRIZ workflow on a gantry crane problem
• Saves a JSON artifact under experiments/ and a Markdown summary
• Runs multi-judge evaluation (default: expert_solution metric)

If Mermaid remote rendering fails (HTTP 502), the notebook falls back to local Graphviz or writes Mermaid source (workflow.mmd) with CLI instructions (mmdc) to render offline.

from triz_agents import create_workflow_app, run_experiment, run_multi_judge_evaluation

# 1) Run an experiment
res = run_experiment(
    model_name="gpt-4o",
    thread_id="gantry-01",
    temperature=0.0,
    out_dir="experiments",
)

# 2) Evaluate with multiple judges
from pathlib import Path
prediction_text = Path("gpt-4o_gantry_*.md").read_text(encoding="utf-8")  # or your artifact path
report = run_multi_judge_evaluation(
    evaluator_models=["o1", "grok-3", "gemini-2.5-pro"],
    input_problem="(the gantry crane problem text...)",
    prediction=prediction_text,
    metrics=["expert_solution"],   # add clarity, coherence, etc.
)

src/triz_agents/tools/triz_tools.py expects:

• src/triz_agents/tools/sources/triz_matrix.xls
• src/triz_agents/tools/sources/triz_principles.json
• src/triz_agents/tools/sources/triz_features.txt

Place your files there or update the paths in triz_tools.py. The contradiction matrix is read via Pandas; column/index names are normalized to lowercase/stripped.

tools/rag_tools.py can build a small TRIZ RAG over PDFs:

• Put your PDFs under ./triz_books/ (or edit the directory).
• Requires chromadb, OpenAIEmbeddings (so set OPENAI_API_KEY).
• The tool function is rag_triz_tool(question: str).

If you don’t plan to use RAG immediately, the imports are safe; the tool won’t run until invoked.

• Run single/batch experiments: experiments.py exposes run_experiment and run_batch.
• Artifacts: JSON (steps + metadata) and a Markdown summary for evaluation pipelines.
• Evaluation: evaluation.py offers multi-judge scoring across metrics (default is just expert_solution). Judges are any models supported by llm.get_llm().

Tip: Use low temperature for judges (the code defaults to deterministic settings).

llm.py maps friendly model names → provider classes:

• OpenAI: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, o1, o3, o4-mini
• Google: gemini-1.5-flash, gemini-1.5-pro, gemini-2.5-flash, gemini-2.5-pro
• DeepSeek: deepseek-chat
• xAI: grok-3, grok-3-mini

You can add your own by extending MODEL_PROVIDER_MAP.

• Mermaid 502 when rendering the graph:

  • The notebook now prefers offline Graphviz (draw_png/draw_svg) and only falls back to Mermaid’s web renderer if needed.
  • Install graphviz + pygraphviz (or pydot) and system binaries.

• TRIZ tools can’t find files:

  • Check paths under src/triz_agents/tools/sources/.

• RAG fails:

  • Ensure OPENAI_API_KEY is set and Chroma is installed.
  • If on Windows without FAISS, Chroma will still work (disk/vector backends).

If you use this project, please cite the paper:

@article{szczepanik2025triz,
  title={TRIZ Agents: A Multi-Agent LLM Approach for TRIZ-Based Innovation},
  author={Szczepanik, Kamil and Chudziak, Jaros{\'L} and others},
  journal={arXiv preprint arXiv:2506.18783},
  year={2025}
}

MIT