| ### Gradio app to run QA on a given website | |
| What does the app do : | |
| - Get input prompt (text or transcribed audio) | |
| - Similarity search in a vector DB, returns top k chunks | |
| - Run llm based on context (using basic prompt - langchain not used yet) | |
| - Return response and metadata (url, scores, text chunks...) | |
| Setup process: | |
| - vector store (FAISS) a given website (sitemap) using langchain | |
| - download a llm | |
| - containarise stt (faster-whisper) due to os incompatibility | |
| - build Gradio app | |
| To build Faster-whisper docker image name "fasterwhisper" | |
| `docker build -t . fasterwhisper` | |
| Create python env 'llama' | |
| ``` | |
| conda create --name="llama" python=3.10 | |
| conda activate llama | |
| pip install -r requirements.txt | |
| ``` | |
| To run the app: | |
| `bash run.sh` | |
| To push to hugging face (including large files) use this command: | |
| `git lfs migrate import --everything` | |
| `git add .; git commit -m "message"; git push -f origin` |