Skip to content

SteadyText

Deterministic text generation and embeddings with zero configuration

PyPI Version Python Versions License

Same input → same output. Every time.

No more flaky tests, unpredictable CLI tools, or inconsistent docs. SteadyText makes AI outputs as reliable as hash functions.

Ever had an AI test fail randomly? Or a CLI tool give different answers each run? SteadyText makes AI outputs reproducible - perfect for testing, tooling, and anywhere you need consistent results.

Powered by Julep

✨ Powered by open-source AI workflows from Julep. ✨

🚀 Quick Start

# Using UV (recommended - 10-100x faster)
curl -LsSf https://astral.sh/uv/install.sh | sh
uv add steadytext

# Or using pip
pip install steadytext

import steadytext

# Deterministic text generation
code = steadytext.generate("implement binary search in Python")
assert "def binary_search" in code  # Always passes!

# Streaming (also deterministic)
for token in steadytext.generate_iter("explain quantum computing"):
    print(token, end="", flush=True)

# Deterministic embeddings
vec = steadytext.embed("Hello world")  # 1024-dim numpy array

# Generate text (pipe syntax)
echo "hello world" | st

# Stream output (default)  
echo "explain recursion" | st

# Wait for complete output
echo "explain recursion" | st --wait

# Get embeddings
echo "machine learning" | st embed

# Start daemon for faster responses
st daemon start

🔧 How It Works

SteadyText achieves determinism via:

  • Fixed seeds: Constant randomness seed (42)
  • Greedy decoding: Always chooses highest-probability token
  • Frecency cache: LRU cache with frequency counting—popular prompts stay cached longer
  • Quantized models: 8-bit quantization ensures identical results across platforms

This means generate("hello") returns the exact same 512 tokens on any machine, every single time.

Daemon Mode (v1.3+)

SteadyText includes a daemon mode that keeps models loaded in memory for instant responses:

  • 160x faster first request: No model loading overhead
  • Persistent cache: Shared across all operations
  • Automatic fallback: Works without daemon if unavailable
  • Zero configuration: Daemon used by default when available
# Start daemon
st daemon start

# Check status
st daemon status

# All operations now use daemon automatically
echo "hello" | st  # Instant response!

FAISS Indexing

Create and search vector indexes for retrieval-augmented generation:

# Create index from documents
st index create *.txt --output docs.faiss

# Search index
st index search docs.faiss "query text" --top-k 5

# Use with generation (automatic with default.faiss)
echo "explain this error" | st --index-file docs.faiss

📦 Installation & Models

Install stable release:

# Using UV (recommended - 10-100x faster)
uv add steadytext

# Or using pip
pip install steadytext

Models

Current models (v1.x):

  • Generation: Qwen3-1.7B-Q8_0.gguf (1.83GB)
  • Embeddings: Qwen3-Embedding-0.6B-Q8_0.gguf (610MB)

Version Stability

Each major version will use a fixed set of models only, so that only forced upgrades from pip will change the models (and the deterministic output)

🎯 Use Cases

Perfect for

  • Testing AI features: Reliable asserts that never flake
  • Deterministic CLI tooling: Consistent outputs for automation
  • Reproducible documentation: Examples that always work
  • Offline/dev/staging environments: No API keys needed
  • Semantic caching and embedding search: Fast similarity matching

Not ideal for

  • Creative or conversational tasks
  • Latest knowledge queries
  • Large-scale chatbot deployments

📋 Examples

Use SteadyText in tests or CLI tools for consistent, reproducible results:

# Testing with reliable assertions
def test_ai_function():
    result = my_ai_function("test input")
    expected = steadytext.generate("expected output for 'test input'")
    assert result == expected  # No flakes!

# CLI tools with consistent outputs
import click

@click.command()
def ai_tool(prompt):
    print(steadytext.generate(prompt))

📂 More examples →

🔍 API Overview

# Text generation
steadytext.generate(prompt: str) -> str
steadytext.generate(prompt, return_logprobs=True)

# Streaming generation
steadytext.generate_iter(prompt: str)

# Embeddings
steadytext.embed(text: str | List[str]) -> np.ndarray

# Model preloading
steadytext.preload_models(verbose=True)

📚 Full API Documentation →

🔧 Configuration

Control caching behavior via environment variables:

# Generation cache (default: 256 entries, 50MB)
export STEADYTEXT_GENERATION_CACHE_CAPACITY=256
export STEADYTEXT_GENERATION_CACHE_MAX_SIZE_MB=50

# Embedding cache (default: 512 entries, 100MB)
export STEADYTEXT_EMBEDDING_CACHE_CAPACITY=512
export STEADYTEXT_EMBEDDING_CACHE_MAX_SIZE_MB=100

🤝 Contributing

Contributions are welcome! See Contributing Guide for guidelines.

📄 License

  • Code: MIT
  • Models: MIT (Qwen3)

Built with ❤️ for developers tired of flaky AI tests.