Build agents with a System 1

Use it where an agent picks among many options (intents, tools, known issues, policy answers) and prompting an LLM with every option stops scaling. MahaBodi is one Rust engine with one JSON API, identical across Rust, Python, Node.js, Java, C#/.NET and Go. It stores memory as a fastmemory topology, finds things again with a hybrid search cascade, and decides with Laya's typed decision models. Every answer says whether it matched, how, and whether to hand off.

Build

Pre-release: build from source. The Laya model is exported once to ONNX (about 1.7 GB), and the export is verified against Laya's PyTorch output.

Quickstart

from mahabodi import Bodi

b = Bodi()
b.ingest(open("kb.md").read(), source="kb")          # memory
r = b.query("refund policy", k=5)                     # r["matched"], r["handoff"], r["stage"], r["hits"]

b.load_laya("models/laya-v2")                         # decisions
a = b.decide({"message": "I was charged twice"},
             {"refund": {"type": "noul", "instructions": "Does `message` ask for a refund?"}})

g = b.decide_with_memory({"question": "Can I get my money back after 30 days?"},
                         {"answer": {"type": "noul", "instructions": "Based on `passage`, is the answer to `question` yes?"}},
                         style="passages")            # grounded in memory, cites its hits

const { Bodi } = require('mahabodi');
const b = new Bodi();
b.ingest(fs.readFileSync('kb.md', 'utf8'), { source: 'kb' });
const r = b.query('refund policy', 5);
b.loadLaya('models/laya-v2');
const a = await b.decide({ message: 'I was charged twice' },
  { refund: { type: 'noul', instructions: 'Does `message` ask for a refund?' } });   // off the event loop

try (Bodi b = new Bodi()) {
    b.ingest(Files.readString(Path.of("kb.md")), "kb");
    String r = b.query("refund policy", 5);            // JSON in, JSON out
    b.loadLaya("models/laya-v2");
    String a = b.decide("{\"message\":\"I was charged twice\"}",
        "{\"refund\":{\"type\":\"noul\",\"instructions\":\"Does `message` ask for a refund?\"}}");
}

using MahaBodi;
using var b = new Bodi();
b.Ingest(File.ReadAllText("kb.md"), source: "kb");
var r = b.Query("refund policy", 5);
b.LoadLaya("models/laya-v2");
var a = await b.DecideAsync(JsonNode.Parse("{\"message\":\"I was charged twice\"}")!,
    new JsonObject { ["refund"] = new JsonObject { ["type"] = "noul", ["instructions"] = "Does `message` ask for a refund?" } });

e, _ := mahabodi.New(nil)
defer e.Close()
e.Ingest(kb, "kb")
r, _ := e.Query("refund policy", 5)
e.LoadLaya("models/laya-v2")
a, _ := e.Decide(`{"message":"I was charged twice"}`,
    `{"refund":{"type":"noul","instructions":"Does `+"`message`"+` ask for a refund?"}}`)

use mahabodi_core::Bodi;
let b = Bodi::from_json("")?;
b.call("ingest", &json!({"text": kb, "source": "kb"}))?;
let r = b.call("query", &json!({"q": "refund policy", "k": 5}))?;
b.call("load_laya", &json!({"dir": "models/laya-v2"}))?;
let a = b.call("decide", &json!({"state": {"message": "I was charged twice"},
    "questions": {"refund": {"type": "noul", "instructions": "Does `message` ask for a refund?"}}}))?;

The snippets are illustrative: exact signatures are in each binding's source and tests (bindings/*).

API methods

Every binding sends these through the same JSON dispatcher, call(method, args).

Result fields

Decision options

Reproduce the benchmarks

Every number in BENCHMARKS.md comes from research/results/*.json, which holds per-item predictions. It is regenerated by research/report.py and tested with exact McNemar.

.venv/bin/python research/bench.py --n 500                 # zero-shot, 6 suites + Laya's mitigations
.venv/bin/python research/bench_massive.py --per-lang 100  # 51 languages
.venv/bin/python research/bench_grounding.py --style-file research/results/tune_grounding.json
.venv/bin/python research/report.py                        # regenerates BENCHMARKS.md

Known limits