MahaBodi (Python)

MahaBodi is a System-1 engine for AI agents, written in Rust. It combines a fastmemory-style topology memory (ingest, a density guard, a query cascade that always returns an answer or a handoff, and hybrid retrieval) with fast typed decisions from Laya models, run through ONNX Runtime.

pip install mahabodi            # memory only
pip install "mahabodi[laya]"    # also installs onnxruntime, for Laya decisions

from mahabodi import Bodi

b = Bodi()
b.ingest(open("kb.md").read(), source="kb")
r = b.query("refund policy")        # r["matched"], r["handoff"], r["stage"], r["hits"]

# Decisions need a Laya model exported to ONNX (see "Models" below)
b.load_laya("models/laya-v2")
b.decide("I was charged twice", {"refund": {"type": "noul", "instructions": "Asks for a refund?"}})

# Learn from labelled cases without retraining (needs the MiniLM embedder export)
b.load_embedder("models/minilm")
b.learn(states, questions, labels)

Every method returns plain dicts and lists; errors raise ValueError.

Models

MahaBodi does not ship model weights. To export Laya's checkpoint and the MiniLM embedder to ONNX, use research/export_onnx.py and research/export_embedder.py in the repository. ONNX Runtime is found through ORT_DYLIB_PATH, or automatically from the onnxruntime package installed in the same Python.

Results

Every claim is measured against Laya on the same machine for each comparison, with seeded samples and exact McNemar tests, and losses are reported as losses. On zero-shot Laya benchmarks MahaBodi wins on many-option and multilingual intent and ties elsewhere (same maths). It loses to a Laya head fine-tuned on SST-5 and to a fully fine-tuned Laya on emotion. See BENCHMARKS.md for all results and caveats.

Status: alpha (0.x). MIT license. Third-party notices are in the repository.