triz-ai 0.7.0

AI-Powered TRIZ Innovation Engine — combines TRIZ methodology with AI and real patent data

triz-ai

AI-powered TRIZ innovation engine — analyze technical problems, classify patents, and discover new inventive principles.

What is this?

triz-ai combines TRIZ (Theory of Inventive Problem Solving) with AI and real patent data. It goes beyond static TRIZ tools by using AI to discover candidate new principles from modern patents, continuing Altshuller's original work.

• Patent-grounded — every suggestion is backed by real patent evidence, with assignee and filing date
• Hybrid search — finds relevant patents using both vector similarity and TRIZ principle/contradiction matching
• Solution directions — generates concrete, actionable solution approaches, not just abstract principles
• 50 engineering parameters — extends Altshuller's 39 with modern domains (security, sustainability, scalability, etc.)
• Evolving principles & parameters — discovers candidate new principles and parameters from modern patents
• Provider-agnostic — works with OpenRouter, Ollama, Anthropic, OpenAI, and 100+ providers via litellm
• Zero infrastructure — local SQLite database with built-in vector search

Installation

pip install triz-ai

Or for development:

# Requires Python 3.12+ and uv
uv sync

Set up your LLM provider API key:

export OPENROUTER_API_KEY="your-key"

Or use a .env file: echo 'OPENROUTER_API_KEY=your-key' > .env

Quick Start

# Analyze a technical problem (works immediately — no setup needed)
triz-ai analyze "How to increase SiC MOSFET switching speed without increasing EMI"
# → Identifies contradiction, recommends TRIZ principles, finds related patents
#   with assignees, and generates concrete solution directions

# For patent-backed examples, ingest patent data
triz-ai ingest data/patents/battery_patents.json

# Discover underused principles in a domain
triz-ai discover --domain "battery technology"

# Run evolution pipeline to find candidate new principles
triz-ai evolve
triz-ai evolve --review  # interactive accept/reject

# Discover candidate new engineering parameters
triz-ai evolve --parameters
triz-ai evolve --parameters --review

# View matrix statistics
triz-ai matrix stats

Commands

Command	Description
analyze	Full TRIZ pipeline: contradiction → matrix → hybrid patent search → solution directions
discover	Find underused principles in a domain and generate patent-grounded ideas
evolve	Discover candidate new TRIZ principles (--parameters for parameters)
ingest	Ingest and auto-classify patents from .txt, .pdf, or .json files
init	Reset the patent database (only needed with --force)
matrix seed	LLM-seed missing matrix cells for params 40-50 (power-user)
matrix stats	Show matrix fill rate and patent observation statistics

All commands support --format text|json|markdown and --model to override the LLM model.

Project Structure

src/triz_ai/
  cli.py              # Typer CLI entry point
  config.py            # Config loading (~/.triz-ai/config.yaml)
  data/                # Static TRIZ JSON data files (principles, parameters, matrix)
  knowledge/           # 40 principles, 50 parameters, contradiction matrix
  engine/              # analyzer, classifier, generator, evaluator
  patents/             # SQLite + sqlite-vec store, ingestion pipeline
  evolution/           # Candidate principle & parameter discovery + review
  llm/                 # litellm wrapper with pydantic validation
tests/                 # 82 unit tests

Configuration

Config lives at ~/.triz-ai/config.yaml:

llm:
  default_model: openrouter/nvidia/nemotron-3-super-120b-a12b:free
  classify_model: openrouter/nvidia/nemotron-3-nano-30b-a3b:free  # smaller model for classification

embeddings:
  model: openrouter/nvidia/llama-nemotron-embed-vl-1b-v2:free
  dimensions: 768

database:
  path: ~/.triz-ai/patents.db

evolution:
  review_threshold: 0.7

Any litellm-supported model string works — just change the model and set the corresponding API key.

Using a custom LLM gateway (e.g., company litellm proxy)

llm:
  default_model: gpt-4o
  classify_model: gpt-4o-mini  # smaller/cheaper model for patent classification during ingest
  api_base: https://llm-proxy.internal/v1
  api_key: your-proxy-token

embeddings:
  model: text-embedding-3-small
  dimensions: 768
  api_base: https://llm-proxy.internal/v1
  api_key: your-proxy-token

You can also override the classify model per-command: triz-ai ingest data/ --classify-model gpt-4o-mini

Development

uv sync                              # Install dependencies
uv run pytest                        # Run tests
uv run pre-commit install            # Install git hooks
uv run pre-commit install --hook-type pre-push

License

MIT