msgmodel 3.2.2

Privacy-focused Python library for interacting with LLM providers (OpenAI, Gemini). Zero-retention middleware with stateless BytesIO-only file uploads.

msgmodel

A unified Python library and CLI for interacting with multiple Large Language Model (LLM) providers with privacy-first, zero-retention design.

Overview

msgmodel provides both a Python library and a command-line interface to interact with two major LLM providers:

• OpenAI (GPT models) — Zero Data Retention enforced
• Google Gemini (Paid tier) — Abuse-monitoring retention only

Privacy Guarantee: All files are processed via in-memory BytesIO objects with base64 inline encoding. No persistent file storage, no server-side uploads, no training data retention (when configured correctly).

Features

• Unified API: Single query() and stream() functions work with all providers
• Library & CLI: Use as a Python module or command-line tool
• Streaming support: Stream responses in real-time
• File attachments: Process images, PDFs, and text files with in-memory BytesIO only
• Flexible configuration: Dataclass-based configs with sensible defaults
• Multiple API key sources: Direct parameter, environment variable, or key file
• Exception-based error handling: Clean errors, no sys.exit() in library code
• Type-safe: Full type hints throughout
• Privacy-first: Mandatory zero-retention enforcement, stateless design, BytesIO-only file transfers

Installation

From PyPI (Recommended)

pip install msgmodel

From Source

# Clone the repository
git clone https://github.com/LeoooDias/msgmodel.git
cd msgmodel

# Install the package
pip install -e .

# Or with development dependencies
pip install -e ".[dev]"

Prerequisites

• Python 3.10 or higher
• API keys from the providers you wish to use

Quick Start

As a Library

from msgmodel import query, stream

# Simple query (uses OPENAI_API_KEY env var)
response = query("openai", "What is Python?")
print(response.text)

# With explicit API key
response = query("gemini", "Hello!", api_key="your-api-key")

# Streaming
for chunk in stream("openai", "Tell me a story"):
    print(chunk, end="", flush=True)

# With file attachment (in-memory BytesIO only)
import io
file_obj = io.BytesIO(your_binary_data)
response = query("gemini", "Describe this image", file_like=file_obj, filename="photo.jpg")

# With custom configuration
from msgmodel import OpenAIConfig

config = OpenAIConfig(model="gpt-4o-mini", temperature=0.7, max_tokens=2000)
response = query("openai", "Write a poem", config=config)

As a CLI

# Basic usage
python -m msgmodel -p openai "What is Python?"

# Using shorthand provider codes
python -m msgmodel -p g "Hello, Gemini!"  # g = gemini
python -m msgmodel -p o "Hello, OpenAI!"  # o = openai

# With streaming
python -m msgmodel -p openai "Tell me a story" --stream

# From a file
python -m msgmodel -p gemini -f prompt.txt

# With system instruction
python -m msgmodel -p openai "Analyze this" -i "You are a data analyst"

# With file attachment (base64 inline)
python -m msgmodel -p gemini "Describe this" -b image.jpg

# Custom parameters
python -m msgmodel -p openai "Hello" -m gpt-4o-mini -t 500 --temperature 0.7

# Get full JSON response instead of just text
python -m msgmodel -p openai "Hello" --json

# Verbose output (shows model, provider, token usage)
python -m msgmodel -p openai "Hello" -v

API Key Configuration

API keys can be provided in three ways (in order of priority):

1. Direct parameter: query("openai", "Hello", api_key="sk-...")
2. Environment variable:
   • OPENAI_API_KEY for OpenAI
   • GEMINI_API_KEY for Gemini
3. Key file in current directory:
   • openai-api.key
   • gemini-api.key

Configuration

Each provider has its own configuration dataclass with sensible defaults:

from msgmodel import OpenAIConfig, GeminiConfig

# OpenAI configuration
openai_config = OpenAIConfig(
    model="gpt-4o",           # Model to use
    temperature=1.0,          # Sampling temperature
    top_p=1.0,                # Nucleus sampling
    max_tokens=1000,          # Max output tokens
)

# Gemini configuration
gemini_config = GeminiConfig(
    model="gemini-2.5-flash",
    temperature=1.0,
    top_p=0.95,
    top_k=40,
    safety_threshold="BLOCK_NONE",
)

Data Retention & Privacy

msgmodel is designed with statelessness as a core principle. Here's what you need to know:

OpenAI (Zero Data Retention)

When using OpenAI:

• What's protected: Input prompts, system instructions, and model responses
• How: The X-OpenAI-No-Store header is automatically added to all Chat Completions requests
• Result: OpenAI does not use these interactions for service improvements or model training
• Persistence: Inputs/outputs are not stored beyond the immediate request-response cycle
• File handling: All files are base64-encoded and embedded inline in prompts — no server-side uploads

Important limitations:

• OpenAI's API logs may retain minimal metadata (timestamps, API version, token counts) for ~30 days for debugging purposes, but not the actual content
• Billing records will still show API usage but not interaction content

Example (ZDR enforced):

from msgmodel import query

response = query("openai", "Sensitive prompt")
# Zero Data Retention is enforced automatically by the X-OpenAI-No-Store header

Google Gemini (Paid Tier Required)

Google Gemini's data retention policy depends on which service tier you use. No API parameter controls this; it's determined by your Google Cloud account configuration.

Paid Services (Google Cloud Billing + Paid Quota)

• What's protected: Data is NOT used for model training or product improvement
• What IS retained: Prompts and responses retained temporarily for abuse detection only (typically 24-72 hours)
• Human review: NO (unless abuse is detected)
• Statelessness: ✅ ACHIEVABLE — within abuse monitoring requirements
• File handling: Base64-encoded inline embedding — no persistent storage

from msgmodel import query

# Paid tier: Data protected from training, used only for abuse monitoring
response = query("gemini", "Sensitive prompt", api_key="your-api-key")

Important: Using Gemini assumes your Google Cloud project has:

1. Cloud Billing account linked
2. Paid API quota enabled (not on free quota tier)

If not enabled, Google will apply unpaid service terms regardless of your code.

Learn more: Google Gemini API Terms — How Google Uses Your Data

Summary Comparison

Provider	Statelessness Achievable	How	Caveat
OpenAI	✅ YES	Automatic ZDR header	Metadata ~30 days
Gemini (Paid)	✅ MOSTLY	Cloud Billing + paid quota	Abuse monitoring ~24-72 hours
Gemini (Unpaid)	❌ NO	No configuration possible	Data retained for training indefinitely

For maximum privacy, use OpenAI with zero-retention (default) or Gemini with paid Cloud Billing.

File Uploads

The BytesIO-Only Approach

All file uploads in msgmodel v3.2.0+ use in-memory BytesIO objects with base64 inline encoding:

import io
from msgmodel import query

# Read file into memory
with open("document.pdf", "rb") as f:
    file_data = f.read()

# Create BytesIO object
file_obj = io.BytesIO(file_data)

# Query with file
response = query(
    "openai",
    "Summarize this document",
    file_like=file_obj,
    filename="document.pdf"  # Enables MIME type detection
)

Why BytesIO only?

• No disk persistence between requests
• No server-side file uploads (Files API not used)
• Better privacy—files never stored on provider servers
• Stateless operation—each request is completely independent

File Size Limits

• OpenAI: ~15-20MB practical limit (base64 overhead + token limits)
• Gemini: ~22MB practical limit (base64 overhead + token limits)

If API returns a size-related error, the file exceeds practical limits for that provider.

Error Handling

The library uses exceptions instead of sys.exit():

from msgmodel import query, MsgModelError, AuthenticationError, APIError

try:
    response = query("openai", "Hello")
except AuthenticationError as e:
    print(f"API key issue: {e}")
except APIError as e:
    print(f"API call failed: {e}")
    print(f"Status code: {e.status_code}")
except MsgModelError as e:
    print(f"General error: {e}")

Response Object

The query() function returns an LLMResponse object:

response = query("openai", "Hello")

print(response.text)          # The generated text
print(response.model)         # Model used (e.g., "gpt-4o")
print(response.provider)      # Provider name (e.g., "openai")
print(response.usage)         # Token usage dict (if available)
print(response.raw_response)  # Complete API response

Project Structure

msgmodel/
├── msgmodel/                    # Python package
│   ├── __init__.py              # Public API exports
│   ├── __main__.py              # CLI entry point
│   ├── core.py                  # Core query/stream functions
│   ├── config.py                # Configuration dataclasses
│   ├── exceptions.py            # Custom exceptions
│   ├── py.typed                 # PEP 561 marker for typed package
│   └── providers/               # Provider implementations
│       ├── __init__.py
│       ├── openai.py
│       └── gemini.py
├── tests/                       # Test suite
│   ├── test_config.py
│   ├── test_core.py
│   └── test_exceptions.py
├── pyproject.toml               # Package configuration
├── LICENSE                      # MIT License
├── MANIFEST.in                  # Distribution manifest
├── requirements.txt             # Dependencies
└── README.md

CLI Usage

After installation, the msgmodel command is available:

# Basic usage
msgmodel -p openai "What is Python?"

# Or using python -m
python -m msgmodel -p openai "What is Python?"

# Provider shortcuts: o=openai, g=gemini
msgmodel -p g "Hello, Gemini!"

# With streaming
msgmodel -p openai "Tell me a story" --stream

# From a file
msgmodel -p gemini -f prompt.txt

Running Tests

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Run with coverage
pytest --cov=msgmodel

Building & Publishing

# Install build tools
pip install build twine

# Build the package
python -m build

# Check the distribution
twine check dist/*

# Upload to PyPI (requires PyPI account)
twine upload dist/*

# Upload to TestPyPI first (recommended)
twine upload --repository testpypi dist/*

License

MIT License - see LICENSE for details.

Author

Leo Dias