asgt-engine 1.0.1

ASGT — CPU-Native Algebraic AI Engine. Zero Float, Pure GF(p) Arithmetic.

ASGT Engine — Algebraic Spectral Generative Theory

CPU-Native AI Engine | Zero Floating Point | Pure Integer Arithmetic on GF(p)

ASGT replaces GPU-bound linear algebra with elliptic curve arithmetic over finite fields. The entire engine runs on the CPU's integer ALU — no VRAM, no float, no external dependencies.

Key Features

• Zero Float/Double: All computation uses uint64_t modular arithmetic on GF(p)
• Elliptic Curve Embeddings: Words → points on E: y² = x³ + ax + b (mod p)
• Hecke Attention: Integer-only attention mechanism replacing matrix multiply
• p-adic Retrieval: Nearest-neighbor search using p-adic valuation distance
• Hensel Training: Exponentially convergent parameter lifting (p-adic Newton)
• Binary Checkpointing: Save/load trained models as compact .asgt files

Installation

pip install asgt-engine

Quick Start

import asgt_engine

# Initialize engine with elliptic curve E(GF(65537))
engine = asgt_engine.ASGTEngine(a=7, b=13, p=65537)

# Train on texts (expands vocab + builds knowledge base)
texts = [
    "The heart pumps blood through the circulatory system",
    "DNA carries the genetic instructions for life",
    "Photosynthesis converts sunlight into chemical energy",
]
engine.train_batch(texts)

# Retrieve relevant knowledge
results = engine.retrieve("What pumps blood?", top_k=3)
for r in results:
    print(f"[sim={r.similarity}] {r.text}")

# Generate answer
answer = engine.generate("What carries genetic information?")
print(answer)  # "DNA carries the genetic instructions for life"

# Save and load model
engine.save_model("my_model.asgt")

engine2 = asgt_engine.ASGTEngine()
engine2.load_model("my_model.asgt")

Architecture

Input Text → Tokenizer → EC Point Embedding → Hecke Attention → p-adic Retrieval → Output
     │            │              │                    │                │
  Strings    Normalize     [n]G on E(GF(p))    Σ[wᵢⱼ]Pⱼ        v_p(x-y)
     │        + Split       try-and-increment   integer weights   integer dist
     ▼            ▼              ▼                    ▼                ▼
  Python       C++ ALU        C++ ALU             C++ ALU          C++ ALU

Performance (Google Colab CPU)

Metric	Value
Encoding speed	44,000+ sentences/sec
Retrieval speed	21,000+ queries/sec
Retrieval accuracy	100% (6/6 benchmark)
Model size	16 bytes per fact
Data types used	uint64_t only

License

MIT