A conversational weather analytics app built around the Open-Meteo API and a local fine-tuned LLM.
Meteo-Chat is an interactive AI-powered weather assistant that combines:
- Real-time and historical meteorological data from the Open-Meteo API, which aggregates ERA5 reanalysis and numerical weather prediction (NWP) models,
- A local Model-Control-Protocol (MCP) server for structured dataset planning and execution, and
- A locally fine-tuned LLM trained on meteorological summaries to produce concise, dataset-bounded insights.
The goal is to provide data-driven, interpretable, and interactive weather summaries—not random text generation.
It can summarize, visualize, and converse about weather variables (e.g., temperature, wind, humidity, rain) for any location and timeframe.
Meteo-Chat runs two components:
To ensure everything works out of the box, please prepare your system as follows.
- Python ≥ 3.9
- Node.js ≥ 18 (recommended: Node 20 LTS)
- npm (installed automatically with Node)
- Git LFS (for LoRA adapter)
Because many users have broken or conflicting Node/npm installs (Homebrew, system Node, legacy Node, etc.),
we strongly recommend installing Node through nvm.
This prevents almost all npm/mcp_server module-not-found errors.
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.0/install.sh | bashRestart your terminal, then verify:
nvm --versionInstall stable Node.js (recommended)
nvm install 20
nvm use 20Check
node -v
npm -vBefore cloning, install Git LFS (one time per system):
brew install git-lfs
git lfs install git clone https://github.com/mitraarka27/Meteo-Chat.git
cd Meteo-Chat
git lfs pull
Use the provided Makefile:
make install
This creates .venv/, upgrades pip, and installs the package in editable mode.
Use the CLI:
source .venv/bin/activate && meteo-chat
This launches: • MCP server → http://127.0.0.1:8787 • Local LLM → http://127.0.0.1:8899 • Streamlit UI → http://localhost:8501
Runs the app with deterministic summaries when the LLM is not available:
meteo-chat-app
Terminal 1 — MCP server
cd mcp_server
npm run dev
Terminal 2 — LLM server
python -m agent.llm_service
Terminal 3 — Streamlit UI
streamlit run apps/streamlit_app/app.py
If the browser doesn’t auto-open, visit: 👉 http://localhost:8501
💡 How to Use
- Enter a place name (e.g., “Kyoto”)
- Choose duration: • Historical → Past climatology (customizable years) • Current → Real-time weather • Forecast → Next 1–16 days
- List variables (comma-separated): temperature, wind, humidity, rain
- Click Submit → Data + visualizations load
- Explore tabs: • Overview → LLM summary, figures, and statistics • Data → Download CSVs
- In the sidebar, chat naturally:
When is max and min temperature seen?→ The app computes or retrieves the answer.
- Queries the Open-Meteo API for historical, current, or forecasted data.
- Converts free-form user variables (like “wind and rain”) into canonical API parameters (like
wind_speed_10m,precipitation). - Produces interactive time-series and box-plots, including rolling mean ± std and diurnal/monthly distributions.
- Generates LLM-based summaries that remain faithful to data (no hallucinations).
- Works even offline or without the LLM, using deterministic statistical fallbacks.
- Provides downloadable CSVs (per variable or combined).
- Features a polished Streamlit UI with a sidebar chat interface for contextual questions.
Meteo-Chat demonstrates the integration of modern AI infrastructure with scientific APIs, combining multiple technical layers:
| Component | Technology / Skill |
|---|---|
| Data Source | Open-Meteo, backed by ERA5 & NWP models |
| Data Orchestration | Custom MCP server for API planning/execution |
| Frontend | Streamlit (custom CSS + responsive Quicksand theme) |
| LLM Integration | Local FastAPI inference server (LoRA fine-tuned on meteorological text) |
| Structured Summaries | Custom JSON-schema interface & deterministic fallbacks |
| Visualization | Matplotlib (time-series, rolling mean, IQR boxplots) |
| Packaging | Python pyproject.toml with CLI launchers (meteo-chat, meteo-chat-app) |
The project also showcases:
- LLM fine-tuning and formatting control
- Schema-aware generation
- API orchestration via MCP
- User-facing data visualization and conversational retrieval
- Three data modes: historical, current, forecast
- Variable aliasing: free-form → canonical Open-Meteo parameters
- Statistical summaries: mean, median, std, IQR, range, variability %, and trends
- Sparse-aware plots: rain/snow shown only when non-zero
- Deterministic fallbacks: e.g., “When is max and min temperature seen?” → computed from diurnal cycle
- Context-aware chat: dataset-bounded responses from local LLM
- CSV exports: combined + per variable
- Modern UI: clean cards, tabs, and sidebar chat with Quicksand typography
⸻
🧩 Behind the Scenes
MCP Server
Handles all interaction with Open-Meteo. Implements /describe_capabilities, /resolve_location, /plan_query, /execute_plan.
Open-Meteo / ERA5
Open-Meteo provides pre-processed ERA5 and NWP model data: • ERA5 (historical): 0.25° reanalysis (ECMWF) • ICON, GFS, GEM, etc. (forecast): multi-model ensemble
Local LLM • Base model: Qwen2.5-1.5B-Instruct • Fine-tuned via LoRA on scientific summaries • Hosted locally via FastAPI • Expects: POST /generate → { "prompt": "...text..." } Returns: { "text": "...answer..." }
When unavailable or inconsistent, deterministic fallbacks (Python-based logic) ensure continuity.
⸻
Challenges: • Maintaining low latency for API + LLM calls • Harmonizing units and variable naming across Open-Meteo datasets • Ensuring model outputs remain factual and schema-constrained
Author: Arka Mitra, Ph.D.
Atmospheric Scientist · Remote Sensing · AI for Climate & Energy Systems
Developed independently, leveraging:
⚖️ License
MIT License © 2025 Arka Mitra You’re free to use, modify, and share this code with attribution.
⸻
⭐ Summary
Meteo-Chat is not a chatbot. It’s a scientific analytics companion—a bridge between numerical weather prediction data and interpretable, LLM-assisted insight. Built end-to-end by a scientist for scientists, data analysts, and anyone curious about the dynamics behind the numbers.
[SCREENSHOTS]
