How ylanv/datalive works
Project type and purpose
• DataLive.ai is a full-stack, desktop-hosted AI assistant for interactive, multimodal data exploration and visualization.
• It combines a Streamlit frontend with a Python backend that (a) profiles and cleans data locally, (b) generates Python code via OpenAI’s GPT-4o, (c) executes that code in a sandbox, (d) renders visualizations, and (e) interprets plots via a HuggingFace Video-LLaMA2B endpoint.
File structure
• Top level:
• main.py (entrypoint)
• pages/… (Streamlit pages: cover.py, data_import.py, data_info.py, main_page.py)
• src/… (modules & tests)
• pyproject.toml, .python-version, .gitignore, uv.lock
• src/model/ contains UI helper functions (ui_utils.py)
• src/openai_caller.py handles prompts and GPT-4o interactions
• src/utils/ contains data-processing modules (analyse.py, cleaning_data.py, data_info_function.py)
• src/tests/ houses unit tests for core functionality
• No persistent database—data lives in memory, loaded from user CSVs.
Key insights from README
• Frontend: Streamlit pages for uploading data, viewing statistics, issuing natural‐language queries, and displaying plots.
• Backend layers:
a. Data profiling & cleaning (pandas, NumPy, Scikit-learn)
b. Code generation (OpenAI GPT-4o)
c. Plot analysis (Video-LLaMA2B on HuggingFace)
d. Local code execution sandbox (Seaborn, Matplotlib)
• Private/local execution emphasis: metadata extraction, local sandbox, no external data upload.