asgt-engine 1.1.0

ASGT Engine: zero-float integer text RAG with FFSC multimodal vision

ASGT Engine

ASGT Engine is a CPU-native, zero-float research engine for integer-only text retrieval and multimodal medical-style retrieval.

Version 1.1.0 combines:

• ASGT text RAG over elliptic-curve semantic lattices on GF(65537)
• FFSC vision encoding over GF(786433) using 2D NTT, Kloosterman coefficients, and algebraic Zernike coefficients over GF(p^2)
• Hecke-style fusion of text and image fingerprints
• Binary save/load for trained symbolic knowledge entries
• A native pybind11 Python extension named asgt_engine

All core similarity and encoding operations use integer arithmetic. No neural network, floating-point vector embedding, or GPU dependency is required by the engine.

Install

pip install asgt-engine

For local development:

python -m pip install build twine pybind11
python -m build

Quick Start

import numpy as np
import asgt_engine

engine = asgt_engine.ASGTEngine()

engine.train_text_batch([
    "The heart pumps blood through the circulatory system",
    "DNA carries genetic instructions for life",
    "Pneumonia may appear as opacity on chest xray",
])

print(engine.retrieve("What pumps blood?", top_k=1)[0].text)

image = np.zeros((256, 256), dtype=np.uint8)
image[96:160, 96:160] = 240

engine.train_vision_batch(
    [image],
    ["Chest xray pattern with focal bright opacity"],
)

matches = engine.retrieve_multimodal(image, top_k=1)
print(matches[0].text)

engine.save_model("medical_symbolic.asgt")

API

engine = asgt_engine.ASGTEngine(a=7, b=13, p=65537)

Text methods:

• train_text_batch(texts)
• train_batch(texts)
• train_step(text)
• retrieve(query, top_k=3)
• generate(prompt, max_tokens=1)

Vision and multimodal methods:

• train_vision_batch(images, texts)
• retrieve_multimodal(image_pixels, top_k=3)

Persistence:

• save_model(path)
• load_model(path)

Stats:

• vocab_size
• num_facts
• num_entries
• train_steps

Image Input

train_vision_batch and retrieve_multimodal accept each image as either:

• a flat list of 65536 integer pixels
• a nested 256 x 256 list
• a NumPy uint8 array with shape (256, 256)
• a NumPy uint8 array with shape (65536,)

Image loading and resizing stay in Python. The C++ engine receives only raw grayscale pixels.

Medical Safety

ASGT Engine is a research and retrieval engine. It is not a medical device and does not provide diagnosis. Production medical SaaS deployments must add clinical validation, data governance, access control, audit logging, PHI protections, and clinician-facing safety review.

License

MIT