FloatChat | AI-Powered Oceanographic Data Discovery

FloatChat is a RAG-based (Retrieval-Augmented Generation) system designed to make complex oceanographic data accessible via natural language. It allows researchers to query the ARGO dataset—a global array of 3,800 free-drifting profiling floats—in plain English, transitioning raw NetCDF scientific files into conversational insights.

system architecture

The system is built on a decoupled, stateless architecture ensuring sub-second response times for metadata queries and robust handling of high-concurrency LLM requests.

technical workflow

01. Context Retrieval: LangChain queries ChromaDB to fetch metadata context (float IDs, locations, parameters) to ground the AI's understanding.

02. Text-to-SQL: An LLM receives the question, schema, and context to generate a high-precision SQL query for the relational database.

03. Data Execution: The system executes the SQL against PostgreSQL (Supabase) and retrieves raw measurement rows.

04. Insight Generation: Raw data is passed back to the LLM for summarization, delivering a human-readable response alongside interactive visualizations.

technical stack

Frontend: React 18, Tailwind CSS, Vercel, Plotly.js, D3.js
Backend: Python, FastAPI, LangChain, Docker, Render
Database: PostgreSQL (Supabase), ChromaDB (Vector Store)
AI: RAG Pipeline, OpenRouter API (LLM Orchestration)
Data: NetCDF (Scientific Oceanographic Format)

prototype demonstration

challenges & feasibility

Challenges

Handling inconsistent schemas in legacy NetCDF files.
Ensuring zero-hallucination SQL generation for scientific accuracy.
Optimizing latency for streaming LLM responses.

Feasibility

Leverages mature orchestration frameworks (LangChain).
Cloud-native deployment enables effortless scaling.
Managed infrastructure reduces operational overhead.