MANN-Engram-Router

license: mit language:

en

zh tags:

medical

multimodal

routing

RAG

siglip

🧠 MANN-Engram-Router Weights

Welcome to the official model hub for MANN-Engram! MANN-Engram is an Edge-Cloud Multimodal Semantic Router designed to shield Cloud VLMs (Vision-Language Models) from "Information Tsunamis" in complex medical and high-noise scenarios.

This repository hosts the custom Skew-Gaussian tensor routing weights required by the MANN-Engram SDK.

🔗 GitHub Repository: MANN-Engram GitHub (Replace with your link)

📦 What's in this Repository?

To keep the installation footprint as light as possible, this repository only contains the custom routing core weights (the Skew-Gaussian compressor).

The foundational multimodal encoder (google/siglip-so400m-patch14-384) is automatically handled and downloaded by the transformers library when you initialize the SDK.

📂 Directory Structure

We provide language-specific routing weights to ensure the highest semantic extraction accuracy:

📁 en/ (English Version - Available Now)

skew_model_v4full_en.pt: The core weights for the mann_engram_en Python SDK.

📁 zh/ (Chinese Version - Coming Soon)

skew_model_v4full_zh.pt: Optimized for Chinese medical contexts and intent extraction.

🚀 Quick Start

1. Install the SDK

Make sure you have installed the MANN-Engram SDK from our GitHub repository:

Clone the GitHub repo and install

pip install -e .

1. Download the Weights

Download the skew_model_v4full_en.pt file from the en/ folder in this repository. We recommend placing it in a weights/ directory at the root of your project.

1. Initialize the Router

from mann_engram_en.router import MANNEngramRouter

Initialize the router (Hardcore Local Privacy Mode)

router = MANNEngramRouter( siglip_model_path="google/siglip-so400m-patch14-384", # Handled via Hugging Face automatically ckpt_path="./weights/skew_model_v4full_en.pt", # The weight file from this repo enable_local_intent=True, # Zero Data Leakage local extraction local_intent_model="Qwen/Qwen2.5-0.5B-Instruct" )

Start purifying your messy clinical data and multimodal files!

results = router.process_session( raw_chat_input="Doctor hurry up! I hit my head on the door yesterday... Check these lab reports.", file_paths=["blood_test.pdf", "history.docx"], image_paths=["head_ct.jpg", "leg_mri.jpg"], top_p=0.85 )

print(f"✅ Successfully extracted core evidence! Stats: {results['stats']}")

🛡️ License & Citation

These weights are released under the MIT License. If you use MANN-Engram in your research or production, please consider citing our GitHub repository.