Nexus Protocol 🧠

The TCP/IP of Latent Space.

Stop wasting tokens. Start sharing minds.

Nexus is a high-speed transport layer for Artificial Intelligence. It replaces slow, expensive natural language communication between Agents with direct, zero-loss Latent Vector Injection.

---

⚡️ Why Nexus?

We are effectively building the Universal Embedding Standard for AI-to-AI communication.

Feature	🐢 The Old Way (JSON/Text)	🚀 The Nexus Way (Vectors)
Medium	English / Text	High-Dimensional Vectors
Cost	High (Generation + Parsing Tokens)	Zero (Hidden State Transfer)
Speed	Slow (Autoregressive Generation)	Instant (Matrix Multiplication)
Loss	High (Semantic drift in parsing)	Lossless (Mathematical Precision)
Privacy	Public (Text is readable)	Obfuscated (Abstract Math)

---

📦 Installation

pip install nexus-protocol

Requires Python 3.9+ and PyTorch.

---

🚀 Quickstart

Transmit a concept from Llama-3 to Mistral without generating a single word of text.

from nexus.client import NexusClient
from nexus.receiver import NexusReceiver
from nexus.protocol import TensorPacket

# 1. Initialize Sender (The Thinker)
sender = NexusClient("meta-llama/Meta-Llama-3-8B")

# 2. Encode a Thought (No text generation happens here)
# "The sky is blue" -> [1, 4096] Float32 Vector
thought_vector = sender.encode_thought("The atmospheric scattering of light...")

# 3. Packetize (TOON Format)
packet = TensorPacket.from_tensor(sender="Llama-3", tensor=thought_vector)

# --- 🌐 NETWORK TRANSMISSION (JSON Serialized) --- 

# 4. Initialize Receiver (The Listener)
receiver = NexusReceiver("mistralai/Mistral-7B-v0.1")

# 5. Inject Thought & Generate
# The model starts generating from the *concept* you injected
output = receiver.receive_thought(packet.model_dump_json())

print(output) 
# Output: "...creates a blue hue visible from the surface."

---

🧠 Core Architecture

Nexus solves the "Tower of Babel" problem in AI. Every model thinks in a different vector space. Nexus provides the Universal Bridge.

1. NexusClient: Extracts the "EOS Token" hidden state from the source model.
2. NexusBridge: A neural adapter that projects the specific model space (e.g., 4096d) into the Universal Nexus Space (2048d).
3. TensorPacket: Wraps the vector in the TOON (Token Oriented Object Notation) format for transport.
4. InverseBridge: Projects the Universal Vector into the Target Model's space.
5. NexusReceiver: Performs Prompt Injection, inserting the vector directly into the model's context window.

---

🗺 Roadmap

• Protocol v1: TOON-compliant packet structure.
• Adapters: Base NexusBridge architecture with Cosine Similarity normalization.
• Genesis Bridge: Pre-trained weights for Llama-3 <-> Mistral <-> GPT-2.
• HuggingFace Hub: Auto-download adapters from nexus-protocol/adapters.

---

🤝 Contributing

We are building the internet for AIs. We need Bridge Builders.

If you have compute, help us train adapters between popular models. See CONTRIBUTING.md for training guides.

License: MIT