sara-engine 0.3.3

A bio-inspired Spiking Neural Network engine with Hugging Face-like pipeline APIs (CPU-only, Backprop-free).

SARA Engine

SARA (Spiking Architecture for Reasoning and Adaptation) Engine is a cutting-edge AI framework that bridges the gap between biological intelligence and modern artificial neural networks.

It provides a highly efficient, event-driven Spiking Neural Network (SNN) core accelerated by Rust, combined with an intuitive PyTorch-like API. SARA goes beyond standard deep learning by natively supporting biological mechanisms such as NeuroFEM, Predictive Coding, and Hippocampal-inspired memory systems.

🧠 Key Features

• High-Performance Event-Driven Core: Rust-based SNN simulation engine that minimizes computational overhead and maximizes simulation speed.
• PyTorch-like API (sara_engine.nn): Build, train, and deploy complex spiking networks using familiar, modular, and declarative syntax.
• Advanced Biologically-Plausible Mechanisms: * Hippocampal Memory System: Long-Term (LTM) and Short-Term (STM) memory supporting Million-Token contexts and SDR (Sparse Distributed Representations).
  • Synaptic Plasticity: Native support for STDP (Spike-Timing-Dependent Plasticity) and Reward-Modulated STDP (R-STDP).
• Spiking LLMs & Transformers: Innovative spike-based attention mechanisms and fully operational Spiking Language Models.
• Function Calling (Agentic SARA): Capable of emitting specific spikes (e.g., <CALC>) to trigger and integrate external Python tools.

🚀 Installation

Ensure you have Python 3.10 or higher and a working Rust toolchain installed.

# Clone the repository
git clone [https://github.com/matsushibadenki/sara-engine-project.git\](https://github.com/matsushibadenki/sara-engine-project.git)
cd sara-engine-project

# Install the package in editable mode (compiles the Rust core automatically)
pip install -e .

(Note: If changes to the core are not reflecting, ensure you re-run pip install -e . to rebuild the Rust extensions.)

💬 CLI Tools & Instruction Tuning

SARA comes with built-in CLI tools to easily interact with and train the engine on custom dialogue data without heavy GPU resources.

Interactive Chat

Start an interactive chat session using the distilled SNN model:

sara-chat --model models/distilled_sara_llm.msgpack

Instruction Tuning (Training)

You can easily fine-tune or override SARA's personality and knowledge using a simple JSONL file.

sara-train data/chat_data.jsonl --model models/distilled_sara_llm.msgpack

JSONL Data Format Example (chat_data.jsonl):

Each line should be a JSON object containing user and sara (or assistant) keys.

{"user": "こんにちは", "sara": "こんにちは！SARAです。何かお手伝いしましょうか？"}
{"user": "SARAって何？", "sara": "私はスパイキングニューラルネットワークで動くローカルAIエンジンです。"}

🌐 Integration Examples

SARA's lightweight CPU inference makes it perfect for integrating into modern web frameworks and bots.

1. FastAPI Integration

Serve SARA via a REST API:

from fastapi import FastAPI
from pydantic import BaseModel
from sara_engine.inference import SaraInference

app = FastAPI()
sara = SaraInference(model_path="models/distilled_sara_llm.msgpack")

class ChatRequest(BaseModel):
message: str

@app.post("/chat")
def chat_endpoint(req: ChatRequest):
sara.reset_buffer()
prompt = f"You: {req.message}\nSARA:"
response = sara.generate(prompt, max_length=100, temperature=0.1)
return {"response": response.strip()}

2. Discord Bot Integration

Build a fast, local Discord bot:

import discord
import os
from sara_engine.inference import SaraInference

intents = discord.Intents.default()
intents.message_content = True
client = discord.Client(intents=intents)
sara = SaraInference(model_path="models/distilled_sara_llm.msgpack")

@client.event
async def on_message(message):
if message.author == client.user:
return

sara.reset\_buffer()  
prompt \= f"You: {message.content}\\nSARA:"  
response \= sara.generate(prompt, max\_length=100, temperature=0.1)  
await message.channel.send(response.strip())

client.run(os.getenv('DISCORD_TOKEN'))

🛠️ Architecture & Modules

• sara_engine.core: The fundamental building blocks, interfacing with the Rust backend.
• sara_engine.nn: High-level PyTorch-like API for model construction.
• sara_engine.models: Pre-built architectures (e.g., SpikingCausalLM, BioTransformer).
• sara_engine.memory: Implementations of SDR, Hippocampus, and Vector Stores.
• sara_engine.agent: Agentic frameworks for MoE and Function Calling.

🗺️ Roadmap & Documentation

To understand the future direction and deep theoretical background of the SARA Engine, check the following documents:

• doc/ROADMAP.md - Short-term development goals.
• doc/SARA_EVOLUTION_ROADMAP.md - Long-term evolutionary roadmap.
• doc/stateful_snn_theory.md - Theoretical background on Stateful SNNs and NeuroFEM.

📄 License

This project is licensed under the MIT License.