edgeimpulse-docs-rag-qwen

Edge Impulse Docs RAG for Qwen

This workspace builds a local Retrieval-Augmented Generation index over the Edge Impulse docs, then queries it with a small Qwen model. It is designed as the training/prototyping repo for an Edge Impulse docs-aware assistant that can later be plugged into the Hugging Face MCP demo or another agent.

Quick Start

cd C:\Users\Eoin\git\edgeimpulse=docs-rag-qwen
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txt
python -m ipykernel install --user --name edgeimpulse-docs-rag-qwen --display-name "Edge Impulse Docs RAG Qwen"

Fetch the docs:

python scripts\fetch_docs.py --source https://docs.edgeimpulse.com/llms.txt

Build the FAISS index. This includes Markdown, text, and PDF files under data/raw_docs, including data/raw_docs/pdfs/*.pdf:

python scripts\build_index.py --docs data\raw_docs --out data\index

Ask a question:

python scripts\ask.py "How do I create an Edge Impulse JWT token?"
python scripts\ask.py "How do I start a training job from the Studio API?"

Run a local API:

python scripts\serve.py --host 127.0.0.1 --port 8080

Then query it:

curl -X POST http://127.0.0.1:8080/ask -H "Content-Type: application/json" -d "{\"question\":\"How do I deploy to Arduino?\"}"

Notebook

Open notebooks/edge_impulse_docs_rag_qwen.ipynb and select the Edge Impulse Docs RAG Qwen kernel.

The notebook runs the same workflow as the scripts:

1. Fetch Edge Impulse docs into data/raw_docs.
2. Chunk and embed Markdown/text docs and PDFs with sentence-transformers/all-MiniLM-L6-v2.
3. Save a FAISS cosine-similarity index in data/index.
4. Retrieve relevant chunks for a question.
5. Generate a grounded answer with Qwen.

Defaults

• Base model: Qwen/Qwen2.5-Coder-0.5B-Instruct
• Optional LoRA adapter: set --adapter eoinedge/edgeai-docs-embedding-qwen1.5-0.5b-instruct
• Embedding model: sentence-transformers/all-MiniLM-L6-v2
• Index: FAISS IndexFlatIP with L2-normalized vectors
• Docs source: https://docs.edgeimpulse.com/llms.txt

Notes

RAG does not train the model weights. It trains/builds the retrieval index. If you later fine-tune a Qwen adapter, keep this RAG layer anyway so the assistant can answer from current docs and PDFs without a new model training run.