RabbitHole: Autonomous Personal Knowledge Agent

The Problem

You've had hundreds of conversations with ChatGPT -- deep dives into reinforcement learning, half-finished side projects, languages you started learning, papers you meant to read. But that knowledge is trapped in a chat log. You never go back. You never follow through.

What RabbitHole Does

1. Ingests your full ChatGPT export (conversations.json)
2. Extracts rabbit holes -- recurring topics you've obsessed over -- using DeepSeek V3.2
3. Autonomously researches each rabbit hole every 6 hours: generates search queries, grounds them with live web results, synthesizes new insights
4. Generates a daily action plan prioritized by urgency, recency, and depth
5. Runs without you -- no prompting, no manual triggers, no intervention

Architecture

conversations.json
       |
       v
  [DeepSeek V3.2 via Akash ML] -- classify into rabbit holes
       |
       v
  [Postgres on Render] -- conversations, rabbit holes, insights, plans
       |
       v
  [Autonomous Agent Loop - every 6h]
       |
       +---> DeepSeek generates search queries
       +---> You.com Search grounds them with real-time web data
       +---> DeepSeek synthesizes insights + urgency scores
       +---> Daily action plan regenerated
       |
       v
  [FastAPI Dashboard on Render] -- view rabbit holes, insights, daily plan

Sponsor Tools Used

Autonomy

• The agent runs on a 6-hour schedule with zero human input
• Each cycle: picks the stalest high-priority rabbit holes, generates fresh queries, searches the live web, synthesizes findings, scores urgency, and rebuilds the daily plan
• New users just upload a file and walk away -- the agent handles everything from classification to first research cycle in the background

Quick Start

git clone https://github.com/kiankyars/rabbithole
cd rabbithole
cp .env.example .env  # fill in API keys
uv sync
uv run python models.py          # create tables
uv run python ingest.py conversations.json  # ingest your history
uv run python agent.py            # run first research cycle
uv run uvicorn main:app --port 8000  # start dashboard

Built With

• deepseek-v3.2-(akash-ml)
• fastapi
• postgresql
• python
• render
• uv
• you.com-search-api