KroxAI-Mini：开源 EAM 证据增强模型原型，开发已停止

KroxAI mini is an open source AI project. Released on 1st November 2025.

DEVELOPMENT DISCONTINUED

KroxAI mini

Untrained simple EAM (Evidence‑Augmented Model) Chatbot with RAG features (e.g for implementing live databases)
Trained (small) version available too.
Alternative to LocalAI/HF Transformers and AI APIs (e.g Claude, OpenAI, ...).

Features (may change as develoment proceeds)

Adjustable server (add different tools for output generation)
RAG
Simple Q/A answering.
Two tokenizers available: HuggingFace(if installed), Simple (work in progress, fallback if no HF tokenizer is installed)

Introducing EAM

Evidence‑Augmented Models (EAMs) are a new class of AI systems that go beyond standard Large Language Models (LLMs).
Instead of generating answers purely from statistical patterns, an EAM is designed to:

Ground every answer in evidence (citations, database fragments, or retrieved documents).
Provide transparency through audit logs, coverage scores, and conflict detection.
Act as an agent by planning and executing tool‑based actions (e.g. retrieval, parsing, calculation).
Stay lightweight and reproducible, so it can run locally on modest hardware.

Formally, while an LLM maps a query Q to an answer A an EAM maps a query Q and evidence E to both an answer A and a step‑by‑step protocol S.

In short:

EAM = “Answers with evidence.”

KroxAI mini is the first open-source prototype of this approach — a compact EAM that demonstrates how evidence‑grounded reasoning can be combined with modular tool use.

Installation

From PyPI (once published)

pip install kroxai-mini

From source

git clone https://github.com/KroxenStudios/KroxAI-Mini.git
cd KroxAI-Mini
pip install -e .

Optional dependencies

Install with server support (FastAPI, uvicorn, BM25):

pip install kroxai-mini[server]

Install with PyTorch support:

pip install kroxai-mini[torch]

Install with HuggingFace Transformers support:

pip install kroxai-mini[transformers]

Install all optional dependencies:

pip install kroxai-mini[all]

Usage

Basic usage - Tokenizer (no dependencies)

from kroxai_mini import SimpleTokenizer

# Initialize tokenizer
tk = SimpleTokenizer()

# Encode text
text = "Hello, world!"
tokens = tk.encode(text, add_bos=True, add_eos=True)
print(f"Tokens: {tokens}")

# Decode tokens
decoded = tk.decode(tokens)
print(f"Decoded: {decoded}")

Advanced usage - Full model (requires numpy)

from kroxai_mini import SimpleTokenizer, TransformerLM
import numpy as np

# Initialize tokenizer and model
tk = SimpleTokenizer()
model = TransformerLM(
    vocab_size=tk.vocab_size, 
    dim=128, 
    n_layers=2, 
    n_heads=4, 
    ff_hidden=256, 
    max_len=128
)

# Generate text
prompt = "Q: What is AI?\nA: "
ids = tk.encode(prompt, add_bos=True)
x = np.array([ids], dtype=np.int64)
y = model.generate(x, max_new_tokens=64)
response = tk.decode(y[0, len(ids):].tolist())
print(response)

Running the server

python examples/server.py

Training a model

python examples/torch_train.py data.json

Interactive chat

python examples/chat.py

Note: KroxAI mini is an example of how EAM works and somehow not ready-to-use.