BanditDB Sandbox

Live Campaigns

Resets nightly at 03:00 UTC ·

Loading...

Try It

Send a prediction and reward the result. The model updates instantly.

1 Select campaign & context

Campaign

Context vector (comma-separated, values in [0,1])

curl Reference

Same calls, from your terminal. Updates as you interact above.

List campaigns

curl -s https://sandbox.banditdb.com/campaigns \
  -H "X-Api-Key: banditdb-demo"

Predict

curl -s -X POST https://sandbox.banditdb.com/predict \
  -H "Content-Type: application/json" \
  -H "X-Api-Key: banditdb-demo" \
  -d '{"campaign_id":"sleep","context":[1.0,0.35,0.50,0.60,0.96]}'

Reward

curl -s -X POST https://sandbox.banditdb.com/reward \
  -H "Content-Type: application/json" \
  -H "X-Api-Key: banditdb-demo" \
  -d '{"interaction_id":"<from predict response>","reward":0.80}'

What the numbers mean

theta_norm

L2 norm of the weight vector θ per arm. Higher = stronger learned signal. A freshly created campaign shows near-zero values; after hundreds of predict→reward cycles the arms diverge as the model learns which context profiles each arm serves best.

Predictions / Rewards

How many times each arm was selected, and how many of those received a reward. A gap means interactions are still in-flight — reward not yet sent. The reward bar shows the fill rate.

Algorithm

LinUCB adds a deterministic UCB bonus: score = θ·x + α·√(x·A⁻¹·x). It always picks the arm with the highest upper confidence bound.

Thompson Sampling samples θ̃ from the posterior N(θ, α²·A⁻¹) — concurrent users naturally diversify arm coverage.

Context vector

Fixed-length array of floats in [0, 1] describing the current situation. The model learns which arms work for which context profiles — personalisation emerges from the predict→reward signal.

This sandbox

Resets nightly at 03:00 UTC. All matrices are wiped and the three demo campaigns are recreated. Play freely — nothing persists. Rate limited to 10 req/s per IP.

★ GitHub 🐍 PyPI 🐳 Docker Hub Apache-2.0