Build your own Visual RAG Application using less than 300 lines of code.
You can talk to your visual documents with LLM.
- 🔥Updates
-
Video search with Chinese🇨🇳 and multi-model support, Llava, Zhipu-GLM4V and Qwen.
python video_search_zh.py --path YOUR_VIDEO_PATH.mp4 --stride 25 --lvm MODEL_NAME
lvm refers to the model we support, could be Zhipu or Qwen, llava by default.
-
Multimodal-RAG to support Word, PPT, CSV, PDF, Email, HTML, Evernote, Video and image.
-
Video search with Qwen-VL to parse the video and Qwen-Tongyi to do RAG. Nomic Embedding supported and Chinese🇨🇳 supported. Super easy to run:
python video_search_qwen.py --path YOUR_VIDEO_PATH.mp4 --stride 25
Stride refers to the frame interval for video parse. For long video parse, you can change stride to big number for higher process speed but less details.
-
- ⭐Features
- Ingest your videos and pictures with Multimodal LLM
- Q&A with LLM about your visual files
- Run locally without compromising your privacy
- Locating the relevant resource with quotation
- Extremely simple with only one python file with no more than 240 lines of code
- 🎥Process
- Parse videos or pictures in the folder into text with LLava, which run locally with ollama
- Ingest the text into vectorDB
- Query it with local LLM.
- 💃Setup
-
Create and activate virtual environment
python -m venv v-rag source v-rag/bin/activate
-
Clone repo and install dependencies
git clone https://github.com/13331112522/v-rag.git cd v-rag python -m pip install -r requirements.txt cp example.env .env
-
Get ready for models
- Put local LLM weights into folder models, supporting any GGUF format, and change the MODEL_PATH in .env for your model path. You can download the weights by visiting Huggingface/theBloke. We use mistral-7b-instruct-v0.1.Q4_K_S.gguf as our LLM for query.
- We currently employed the HuggingfaceEmbedding, but you can change it to local embedding like GPT4ALLEmbedding by changing the EMBEDDINGS_MODEL_NAME in .env.
- Run MLLM. We employ the latest llava 1.6 for image and video parsing.
ollama run llava
-
Environment variables setting
- Change the environment variables according to your needs in .env.
- Replace with the actual path to your FFmpeg executable in os.environ["IMAGEIO_FFMPEG_EXE"] = "/path/to/ffmpeg" to leverage the FFmpeg backend.
-
Run
Put all the images and videos you want to talk with into the folder source. Run following command:
python v-rag.py --path ./source --stride 25
Path refers to the folder which contains all the images and videos you want to retrieve, and stride refers to the frame interval for video parse. For long video parse, you can change stride to big number for higher process speed but less details.
It will generate the folder source_documents as the storage of parsed text and faiss_index as the vectorDB. If the two folders already exist, it will start query directly.
-
- 🍺Discussion
- Low parsing speed.
- Hallucination.
- Safety constraint
- Examples for parsed text:
- ☕To-do List
- WebUI
- JSON support rather than txt
- Video playback with the query result.
- Evaluation on open Dataset
- MultiModal RAG
- Docker support
- ❤️Acknowledgement
- llava 1.6
- PrivateGPT
- ollama
- langchain
- Llama.cpp