This repository contains TVCache and an end-to-end training example that demonstrates its use. TVCache speeds up RL training of tool-using agents by caching tool execution results in a Tool Call Graph (TCG). TCG is a prefix tree of tool call sequences that enables reuse across rollouts and epochs.

├── tvcache/            # TVCache server and client library
│   ├── server/         # HTTP server that maintains the TCG
│   └── client/         # Python client (tvclient) for integrating TVCache into training loops
├── train/              # Example: Video QA agent trained with Tinker API
└── video-agent-tools/  # Video sandbox server (tool execution backend for the example)

cd tvcache/server
uv sync
uv run tvcache_server.py    # http://localhost:8000

See the server README for API details and configuration.

The example uses a video analysis sandbox. See the video-agent-tools README for setup.

cd train
uv sync
uv pip install -e ../tvcache/client   # install tvclient

Set your Tinker API key in run.sh, then:

./run.sh train_with_tvcache.py

See the training README for dataset preparation and configuration. See the integration guide for how to integrate tvclient into your own training loop.