LLM Function Calling Patterns

Production-ready implementations of advanced LLM function calling patterns in TypeScript and Python. This library provides battle-tested implementations of ReAct (Reasoning and Acting) and ReWOO (Reasoning Without Observation) patterns for building sophisticated AI agents.

Features

• Dual Language Support: Full implementations in both TypeScript (Node.js) and Python
• Multiple LLM Providers: OpenAI, Anthropic, Groq, Together AI, DeepSeek, Cerebras, Ollama
• Shared Configuration: Common environment variables work across both implementations
• Modern Stack:
  • Node.js 22+, TypeScript 5, Vite, ESM modules
  • Python 3.12+, Pydantic, uv package manager
• Production Ready: Error handling, retries, content moderation, logging

Project Structure

.
├── .env.example          # Shared environment configuration
├── README.md             # This file
├── node/                 # TypeScript/Node.js implementation
│   ├── package.json
│   ├── README.md
│   └── src/
└── python/               # Python implementation
    ├── pyproject.toml
    ├── README.md
    └── src/

Quick Start

Node.js (TypeScript)

cd node
npm install
cp ../.env.example .env
# Edit .env with your API keys
npm run react-cli    # Run ReAct CLI
npm run rewoo-cli    # Run ReWOO CLI

Python

cd python
# Using uv (recommended)
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"
cp ../.env.example .env
# Edit .env with your API keys
uv run react-cli    # Run ReAct CLI
uv run rewoo-cli    # Run ReWOO CLI

Shared Configuration

Both implementations use the same environment variables. Copy .env.example to .env in either the node/ or python/ directory.

Required Environment Variables

Variable	Description	Example
LLM_PROVIDER	Provider name	openai, anthropic, groq
LLM_MODEL	Model identifier	gpt-4o, claude-3-5-sonnet-20241022
OPENAI_API_KEY	OpenAI API key	sk-...
ANTHROPIC_API_KEY	Anthropic API key	sk-ant-...
GROQ_API_KEY	Groq API key	gsk_...

Optional Environment Variables

Variable	Default	Description
LLM_BASE_URL	Provider default	Custom API endpoint
LLM_TIMEOUT_MS	30000	Request timeout in milliseconds
LLM_MAX_RETRIES	3	Maximum retry attempts
TEMPERATURE	0.7	Sampling temperature (0-2)
MAX_TOKENS	Provider default	Maximum completion tokens
TAVILY_API_KEY	-	For web search functionality

External Services

Variable	Description
TAVILY_API_KEY	Web search API key
DOJO_API_KEY	Document library API key
DOJO_API_BASE_URL	Document library base URL
DOJO_API_LIBRARY_UUID	Document library UUID
DATABASE_URL	PostgreSQL connection string
POSTGRES_HOST	PostgreSQL host
POSTGRES_PORT	PostgreSQL port
POSTGRES_DB	PostgreSQL database name
POSTGRES_USER	PostgreSQL username
POSTGRES_PASSWORD	PostgreSQL password

Architecture

ReAct Pattern

The ReAct (Reasoning and Acting) pattern interleaves reasoning traces with actions:

User Question → Thought → Action → Observation → Thought → ... → Final Answer

Key Components:

• Agent orchestrates the reasoning loop
• Tools execute actions and return observations
• Message handler manages conversation history
• Iteration control prevents infinite loops

ReWOO Pattern

The ReWOO (Reasoning Without Observation) pattern separates planning from execution:

Task → Plan (all steps) → Execute Steps → Synthesize Solution

Key Components:

• Planner creates execution plans with evidence variables (#E1, #E2, etc.)
• Worker executes each step using available tools
• Solver synthesizes final solution from collected evidence
• Event bus manages reactive event flow

Supported Providers

Provider	Models	Function Calling	Vision
OpenAI	GPT-4o, GPT-4o-mini, GPT-4 Turbo	Yes	Yes
Anthropic	Claude 3.5 Sonnet, Claude 3 Opus/Haiku	Yes	Yes
Groq	Llama 3.3 70B, Mixtral 8x7B	Yes	No
Together AI	Llama 3.1, Qwen 2.5	Yes	No
DeepSeek	DeepSeek Chat/Coder	Yes	No
Cerebras	Llama 3.1 70B/8B	Yes	No
Ollama	Any local model	Yes	Varies

Usage Examples

Node.js (TypeScript)

import { ReActAgent, loadReActConfig } from 'llm-function-calling-patterns';

const config = loadReActConfig();
const agent = new ReActAgent(config.ai, config.tools, config.maxIterations);

const answer = await agent.answer('What is 15% of 250?');
console.log(answer);

Python

import asyncio
from llm_patterns import ReActAgent, load_react_config
from llm_patterns.react.tools import CalculatorTool

async def main():
    config = load_react_config()
    agent = ReActAgent(
        config=config.ai.resolve(),
        tools=[CalculatorTool()],
    )
    answer = await agent.answer("What is 15% of 250?")
    print(answer)

asyncio.run(main())

Development

Node.js

cd node
npm install
npm run dev           # Development server
npm test              # Run tests
npm run test:coverage # Coverage report
npm run build         # Production build

Python

cd python
uv pip install -e ".[dev]"
pytest                # Run tests
pytest --cov          # Coverage report
ruff check .          # Linting
mypy src              # Type checking

Documentation

• Node.js README - Detailed Node.js/TypeScript documentation
• Python README - Detailed Python documentation

References

• ReAct: Synergizing Reasoning and Acting in Language Models
• ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models

License

MIT License - see LICENSE for details.