Hybrid analysis of nested (ragged) HEP data: flatten ROOT events into MonetDB tables, push selections down to SQL, then rebuild the nesting in Awkward Array for the physics. An LLM agent can drive the same pipeline through MCP.
python -m pip install awkward-monetizer # core
python -m pip install "awkward-monetizer[agent]" # + LLM agent and MCP serverPython 3.10+. The MonetDB server is installed separately (brew install monetdb
on macOS). For development (pip or pixi), see
docs/usage.md.
awkward-monetizer make-nano # synthetic NanoAOD file
awkward-monetizer adl # IRIS-HEP ADL Q1–Q8 on it, no database neededWith a MonetDB server running, load real CMS Open Data:
monetdb create hep && monetdb release hep
awkward-monetizer ingest <NanoAOD file>.root --dataset nanoaod --database hep \
--create-schema --step-size 50000 --entry-stop 100000An LLM agent gets six tools: schema discovery, read-only SQL, Awkward code on events selected in SQL, submit, plot, and (over MCP) a check against the ADL reference. To use it from Claude Code:
claude mcp add -s user awkward-monetizer -- awkward-monetizer agent mcp --database hepIt speaks MCP over stdio (started by the client, as above) or HTTP
(awkward-monetizer agent mcp --database hep --transport http, local-only by
default).
On 100,000 CMS Open Data events (DoubleMuon Run2016H, record 30555), Claude answered all eight ADL questions correctly on the first attempt, doing cuts in SQL and combinatorics in Awkward. On the full record, 48.9 million events streamed from EOS in 22 minutes, all eight reproduce the reference exactly on a laptop, the SQL ones in under 3 s each. See docs/agent.md.
- docs/usage.md: development setup, datasets, ingestion, sharding, benchmarks, notebooks, tests
- docs/agent.md: the agent, its tools, evaluation and demo
- CONTRIBUTING.md: how to contribute and how releases are made
A finished prototype in maintenance mode: bug fixes and releases continue, no new features are planned. Arrow transport, electrons and b-tag are open for anyone who wants to take them on.