Chat Agent

An AI agent with tool use, retrieval over uploaded documents, and a persistent E2B Python sandbox for code execution. Ships with both a terminal CLI and a Flask server that streams agent events over Server-Sent Events.

Features

• LLM-driven agent loop built on LangChain (defaults to Google Gemini).
• Persistent E2B sandbox for running Python, shell commands, and generated code with state preserved across turns.
• Document RAG — upload PDFs, DOCX, code, CSV, etc. Files are chunked, embedded with sentence-transformers, indexed with FAISS + BM25, and stored per-session under a namespaced index.
• Web search via Tavily.
• Session persistence — chat history, uploaded files, and auto-generated chat titles are stored in SQLite/Postgres via SQLAlchemy.
• Two entry points: a CLI (main.py) and a streaming HTTP server (server.py).

Requirements

• Python 3.10+
• An E2B account (for sandboxed code execution)
• A Google AI Studio key (for Gemini)
• A Tavily key (for the search tool)

Setup

python -m venv venv
source venv/bin/activate            # Windows: venv\Scripts\activate
pip install -r requirements.txt

Copy the example environment file and fill in your keys:

cp .env.example .env

Then edit .env and set GOOGLE_API_KEY, E2B_API_KEY, and TAVILY_API_KEY. See config/settings.py for the full list of tunables (model name, temperature, token limits, alternate LLM providers, etc.).

Running

CLI

python main.py

In-session commands:

Command	Description
/upload <path>	Upload and index a file (PDF, CSV, code, etc.)
exit / quit	End the session
Ctrl+C	Graceful shutdown

HTTP server

python server.py

Listens on http://0.0.0.0:5000. Endpoints:

Method	Route	Purpose
GET	/sessions	List saved sessions (id, title, created_at)
POST	/upload	Upload files for a session (multipart form)
POST	/chat	Stream a chat turn as SSE agent events
DELETE	/session/<plan_id>	Tear down a session and its sandbox

/chat accepts JSON of the form {"plan_id": "...", "message": "...", "chat_history": [...]} and streams events with type in {status, tool, observation, result, error}.

Project layout

app/
  agent/        # LLM engine + system prompt
  api/          # Pydantic schema for streamed agent events
  sandbox/      # E2B sandbox handler
  tool/         # Tool implementations (search, code, file, terminal)
  utils/        # DB models, logger, title generator
config/         # Pydantic settings loaded from .env
persistent_storage/  # Per-session uploads + FAISS indexes (gitignored)
main.py         # CLI entry point
server.py       # Flask + SSE entry point

Notes

• All runtime configuration lives in config/settings.py and is loaded from .env via pydantic-settings.
• If you supply your own E2B_TEMPLATE_ID, the template must be based on the E2B code-interpreter runtime.
• persistent_storage/ and .env are gitignored — keep them that way.