MLUX-Reactly Agent Framework

A ReAct inspired Agent developed from scratch using Ollama. This agent uses task decomposition and uses provided tools.

Installation

A virtual environment is recommended:

python3 -m venv .venv
source .venv/bin/activate

From PIP

pip install mlux-reactly

From Source

pip install git+https://github.com/d-volution/mlux-reactly.git

Setting up Ollama

You also have to setup Ollama. See the README of Ollama for that.

The agent uses by default the Ollama model qwen2.5:7b-instruct-q8_0, but can be configured via the llm keyword argument. The demo RAG-tool included within this repository uses the embedding model nomic-embed-text.

ollama pull qwen2.5:7b-instruct-q8_0
ollama pull nomic-embed-text

Usage

from typing import Annotated
from mlux_reactly import ReactlyAgent, LLM

def count_substr(a: Annotated[str, "Some string"], b: Annotated[str, "substring to be counted"]) -> int:
    """This tool calculates how often string b occures in string a"""
    if not b:
        return 0
    return sum(1 for i in range(len(a) - len(b) + 1) if a[i:i+len(b)] == b)

agent = ReactlyAgent(tools=[count_substr], llm=LLM("qwen2.5:7b-instruct-q8_0"))

answer = agent.query("How many times does the letter l occure in 'artificial general intelligence'?")
print(f"agent answer: {answer}")

---

How to trace

Use LIVE_VERBOSE for printing the most important events live:

from mlux_reactly.tracer import TestTracer, LIVE_VERBOSE
# ...
agent = ReactlyAgent(tracer=TestTracer(live_format=LIVE_VERBOSE))

Or to print only toolruns and results without highlight colours live:

from mlux_reactly.tracer import TestTracer, FormatConfig
# ...
agent = ReactlyAgent(tracer=TestTracer(live_format=FormatConfig(show=['toolrun', 'result'], colored=False)))

These are the existing events:

key	description
query	an agent.query() call
task	subtask the query is splitted into
stage	a stage doing a LLM call with one specific prompt. If it fails, it might retry resulting in multiple LLM calls.
llmcall	a single call of the LLM. Only shows up in trace if not in compact mode
toolrun	The invocation of a user-supplied tool
root	Root node of tracer
failed	Some step/operation that failed
result	The operation of the parent event finished with some result

When running from Git Repo

Local Install

pip install -r requirements.txt

Run simple chatbot

python3 test/chat.py

Evaluation

To evaluate the agent, you can use eval.py like this:

python3 test/eval.py -agents <agent names> -tests <test names with test params>

For example, use

python3 test/eval.py -agents reactly llama-react -tests hotpot/train:100:3

to run the 'hotpot' test with the examples from index 100 to 103 on both the Reactly and the Llama-ReAct agents.