TRIDENT — 3-Head Tiny LM + P2P RAG + MCP

~1M parameter transformer with three specialist heads: Code, Math, Research. Runs on CPU (Termux/iPad/laptop). Knowledge retrieval via P2P WebRTC — no cloud needed.

Files

• model.py — Trident architecture (backbone + 3 heads + RAG gate)
• rag.py — Python chunk store + P2P RAG node (aiortc)
• rag_client.js — Browser P2P RAG client (WebRTC DataChannel)
• signal_server.py— Minimal WebSocket signaling (handshake only, ~50 lines)
• mcp_server.py — FastMCP server exposing all tools to Claude/any LLM
• train.py — Training loop (toy data → real data)

Quick Start

# Install
pip install -r requirements.txt

# Test model architecture
python model.py

# Train on toy data (CPU, ~1 min)
python train.py

# Run MCP server (connect to Claude Desktop)
python mcp_server.py

# Run signaling server for P2P
python signal_server.py

P2P RAG Flow

Device A (your phone)
  └─ TRIDENT_P2P.query("fibonacci")
       └─ WebRTC DataChannel → Device B (iPad)
            └─ Device B searches local chunks
            └─ Returns top-K matches
       └─ Device A merges local + peer results
       └─ Top-K embeddings → RAGFusionGate → model generates

MCP Tools

Tool	What it does
trident_generate	Generate text, pick head, optionally RAG
trident_add_chunk	Add knowledge to RAG store
trident_search_rag	Search chunks without generating
trident_router	Predict which head fits a query
trident_list_chunks	List all stored knowledge chunks

Architecture

Input → [Embedding + PositionalEncoding]
      → [Backbone: 4x TransformerBlock]  ← shared 
      → [HeadRouter]  → softmax weights
      → ┌─────────────────────────────┐
        │ RAGFusionGate (cross-attn)  │  ← per head
        │ 1x TransformerBlock         │
        │ LM Head → logits            │
        └─────────────────────────────┘ × 3 heads
      → weighted ensemble OR forced single head