TouchDesigner RAG

A local Retrieval-Augmented Generation (RAG) system for answering questions about TouchDesigner using wiki pages, tutorials, and other structured or semi-structured content. Powered by FAISS and local LLMs via Ollama.

---

Installation

• Install the required environment.
• Make sure Ollama is installed and running.

---

Files

• model/ - Precompiled FAISS index and metadata files. Latest precompiled FAISS index is available at: huggingface.co/imjj/touchdesigner-rag-wiki-index
• scripts/ - All Python scripts for preprocessing, embedding, and launching the app.
• td/ - Contains components for use inside TouchDesigner.

---

Preprocess Raw HTML (wiki_cleaner.py)

This script is used to chunk TouchDesigner wiki HTML pages, typically found under:

Derivative\TouchDesigner\Samples\Learn\OfflineHelp\https.docs.derivative.ca

python wiki_cleaner.py \
  --input_dir data/raw_html \
  --output_dir data/cleaned/docs \
  --jsonl data/cleaned/chunks.jsonl

---

Build Embeddings + FAISS Index (build_embeddings.py)

This script is used to generate a FAISS vector index from the JSONL chunks using a specified embedding model.

python build_embeddings.py \
  --chunks_path data/cleaned/chunks.jsonl \
  --index_path model/td_index.faiss \
  --metadata_path model/td_metadata.json \
  --embedding_model all-mpnet-base-v2

---

Launch the Gradio App (app.py)

This script launches the Gradio web interface for querying the RAG system.

cd scripts
python app.py

Then visit http://127.0.0.1:7860 in your browser.

Alternatively, you can try a simplified FAISS retrieval demo online (semantic search only, without LLM reasoning):
TouchDesigner Wiki Search (Hugging Face Space)

---

Use Inside TouchDesigner (Optional)

After launching the Gradio app, you can use the .tox component by dragging td_rag_v1.0.tox into your TouchDesigner network to:

• input your question directly inside TD
• trigger the RAG-based LLM to respond automatically
• view answers and references without leaving TouchDesigner

td_rag_v1.0.tox

• Make sure the Gradio app is running in the background.
• Press Reset to retrieve the available models from the backend before running.
• Enter your question in the Query field and press Run to fetch the result.

---

Notes and Limitations

Precompiled Data

The included FAISS index and metadata are built exclusively from the offline TouchDesigner wiki (`Samples\Learn\OfflineHelp`). All documents were automatically extracted and chunked without thorough manual review. You're encouraged to refine the dataset or expand it by incorporating other sources (e.g., forums, tutorials). A more curated version may be released in future updates.

LLM Prompt Design

The current LLM system prompt is a basic template aimed at guiding the model's reasoning. Feel free to adapt or enhance it to better fit your technical needs or creative workflow.

---

Requirements

• Python 3.10+
• Ollama installed and running locally

---

License

This project is licensed under the MIT License.

Notes:

• The precompiled index was generated from the offline TouchDesigner documentation located in Samples/Learn/OfflineHelp. This content remains the property of Derivative Inc. and is used here for educational and personal purposes only.
• The included .tox component was created using the TouchDesigner Non-Commercial Edition. Users are responsible for ensuring compliance with their own TouchDesigner license (Commercial, Educational, or Non-Commercial).
• This project uses open-source libraries including FAISS, SentenceTransformers, and Gradio. It also integrates with large language models (LLMs) served via Ollama, which are subject to their respective licenses.