ChatPilot is an agentic workstation featuring self-correcting decomposition RAG with visual PDF highlighting and deep reasoning. It automates secure containerized data analysis and deep scholarly research, supporting OpenAI, Claude, Gemini, DeepSeek, and Ollama models.

ChatPilot: Intelligent Agentic RAG & Autonomous Research Engine

Demo video: Watch demo video

ChatPilot is a tool-augmented Agentic DeepSeek that goes beyond simple chat. It utilizes a single AI agent to interact with your local files, conduct deep web research, and perform autonomous data analysis through Python code execution.

Demo Frontend: ChatPilot Frontend

Key Capabilities

Document Intelligence & Agentic Retrieval

Powered by Docling, ChatPilot supports both digital and scanned documents with layout-aware parsing and OCR.

• Multi-keyword and clause-level retrieval
• Dynamic relevance re-evaluation by the agent
• Deep context retrieval when shallow chunks are insufficient
• Coordinate-based highlighting of retrieved content directly in PDFs

Scholarly Research & Visual RAG

ChatPilot includes a research-grade scholarly retrieval pipeline focused on traceability.

• Integrated with OpenAlex (250M+ scholarly works)
• Natural language queries over academic literature
• Query decomposition RAG for higher recall and precision
• Retrieved results are reranked using semantic similarity

Source Highlighting

• All sources used by the LLM are returned with each response
• Each retrieved chunk is mapped to its exact location in the source document
• Referenced text is highlighted directly in the PDF viewer
• Multiple sources are surfaced and highlighted independently

This allows direct verification of claims without manually searching papers.

Autonomous Data Analysis

ChatPilot functions as a data analyst agent for structured datasets.

• Upload large CSV or Excel files (no file-size limits)
• The LLM never accesses row-level data
• Python code is generated using schema-level context only
• Code executes inside an isolated Docker sandbox

Agent loop:

• Plan → Execute → E