About this Agentic Data Analyst

Analyse structured data in natural language — no SQL required.

This is an agentic data analyst — a tool that lets you explore and analyse structured datasets through a natural language conversation, without writing a single line of SQL. It is being built in four phases, as outlined below. The current version runs on Claude Sonnet, Anthropic's state-of-the-art language model. Support for local open-source models is planned for a later phase.

Upload one or more CSV files and ask business questions in plain English. The agent translates your question into SQL, executes it against your dataset, and returns a clear, grounded answer — all results are backed by actual query execution, not model inference.

Across multiple analyses, the tool builds up a memory of past findings. Built on a LangGraph agent with a two-tier architecture, past analyses are stored in a Chroma vector database and retrieved via RAG (Retrieval-Augmented Generation), so relevant context is injected into each new question without passing the full conversation history on every API call.

The analysis phase is instrumented with LangSmith, providing full observability into every agent run — including the SQL queries generated, tool calls made, and token usage per step.

Phase 1 — Live ✓

• Claude API Integration
• Schema Confirmation & NL→SQL Analysis Loop
• RAG Memory: vector store for cross-analysis context

Phase 2 — Live ✓

• AWS Cloud Deployment
• Docker Containerisation
• CI/CD Pipeline

Phase 3 — In Progress

• Monitoring & Observability: LangSmith tracing ✓
• Kaggle Integration: load datasets directly by URL
• User Accounts & persistent session storage

Phase 4 — Planned

• Local Model Support: run without an API key via Ollama
• Data Transformation: natural language to pandas operations

Simplified — actual flow includes conditional branches and conversational loops. See README for details.

Important Notes

API key required. This tool uses the Anthropic Claude API. You need to bring your own key from console.anthropic.com. Your key is transmitted to the server only to make API calls and is wiped from memory at the end of your session — it is never logged or stored to disk.

Prefer to run locally? Clone the repo at github.com/Wilsbert12/agentic-data-analyst and follow the setup instructions in the README.

How to Use

Six steps from dataset to insight.

Demo

See the agent in action.

1. Setup

1. Navigate to the Try it section
2. Clear memory from earlier runs (optional)
3. Enter your Claude API key
4. Press Start Session
5. Upload your dataset (CSV files)
6. Upload additional context files — data cards, schema images (optional)
7. Press Analyse Dataset

2. Full Walkthrough

Click on the video to see a complete interaction with the agentic data analyst.

1. Setup — the API key and files are uploaded as described in step 1
2. Schema confirmation — the agent analyses the uploaded dataset and summarises tables, column types, and key relationships
3. Clarifying questions — ambiguities are flagged and resolved before any analysis begins
4. Schema confirmed — confirmed rules are saved and inform every subsequent query
5. Business question — a natural language question is submitted; the agent asks a clarifying question before moving to the analysis
6. Result — the agent runs SQL against the database and returns a grounded answer
7. Refine or move on — the user is not satisfied with the first result and asks a follow-up question to refine it; after the second analysis, the user confirms the result and moves to the next question
8. Memory — confirmed analyses are summarised and stored; past context is retrieved automatically for future questions

Try it

Upload your dataset and start asking questions.