alloy-ai 0.1.1

Alloy (Python): Python for logic. English for intelligence.

Alloy (Python): Python for logic. English for intelligence.

License: MIT

This repository contains an early scaffold of the Alloy library per alloy-spec-v1.md.

Quick start

• Install (all providers): pip install 'alloy-ai[providers]'
• Or minimal (OpenAI only): pip install alloy-ai
• Create .env with OPENAI_API_KEY=...
• Use the API:

from dataclasses import dataclass
from dotenv import load_dotenv
from alloy import command, ask, configure

load_dotenv()
# Optional: configure() — default model is `gpt-5-mini` if omitted
# configure(model="gpt-5-mini", temperature=0.7)

@command(output=float)
def ExtractPrice(text: str) -> str:
    """Extract price from text."""
    return f"Extract the price (number only) from: {text}"

print(ExtractPrice("This item costs $49.99."))
print(ask("Say hi"))

Notes

• OpenAI backend is implemented for sync/async/streaming.
• Streaming with tools is not yet supported.
• For structured outputs, Alloy attempts to use OpenAI structured responses (JSON schema). If unavailable, the model may still return JSON, which Alloy parses best-effort.
• Configuration defaults: Alloy uses model=gpt-5-mini if configure(...) is not called. You can also set process environment variables instead of a .env file:
  • ALLOY_MODEL, ALLOY_TEMPERATURE, ALLOY_MAX_TOKENS, ALLOY_SYSTEM/ALLOY_DEFAULT_SYSTEM, ALLOY_RETRY.
  • Example: export ALLOY_MODEL=gpt-4o then run your script.

Examples

• See examples/basic_usage.py and examples/tools_demo.py (tools + contracts).

Offline mode (dev only)

• To run examples without network/API keys, set ALLOY_BACKEND=fake.
• Example: ALLOY_BACKEND=fake python examples/basic_usage.py

Config precedence

• Defaults: model=gpt-5-mini, max_tool_turns=2.
• Process env (ALLOY_*) overrides defaults.
• Context/use_config and configure(...) override env/defaults.
• Per-call overrides (e.g., ask(..., model=...)) override everything above.

Make targets

• make setup — install dev deps and package in editable mode.
• make test, make lint, make typecheck — CI-like checks.
• make examples — runs examples/basic_usage.py and examples/tools_demo.py.
  • Tip: ALLOY_BACKEND=fake make examples to run offline.

Troubleshooting

• API key: Ensure OPENAI_API_KEY is set (process env or .env).
• Model choice: Prefer gpt-5-mini for fastest latency; switch via configure(model=...) or ALLOY_MODEL.
• Timeouts/slow runs: Reduce max_tokens, lower temperature, prefer smaller models, and cap tool loops.
• Tool loops: Alloy caps tool iterations by default (max_tool_turns=2). Adjust via configure(max_tool_turns=1) or env ALLOY_MAX_TOOL_TURNS.
• Rate limits (429): Shorten prompts/outputs, add retries with backoff, or use lower-throughput settings.

Integration tests

• OpenAI: Set OPENAI_API_KEY (and optionally ALLOY_IT_MODEL, default gpt-5-mini). Run pytest -q — OpenAI integration tests auto-enable.
• Anthropic: Set ANTHROPIC_API_KEY and ALLOY_IT_MODEL=claude-3.5-sonnet (or another Claude). Run pytest -q — Anthropic integration tests auto-enable.
• Gemini: Set GOOGLE_API_KEY and ALLOY_IT_MODEL=gemini-1.5-pro (or another Gemini). Run pytest -q — Gemini integration tests auto-enable.
  • SDK note: Gemini support uses google-genai (GA).

How to run locally

• Install providers bundle: pip install 'alloy-ai[providers]'
• Create .env with OPENAI_API_KEY=...
• Option A (no install):
  • python examples/basic_usage.py
  • python examples/tools_demo.py
  • (examples add src/ to sys.path for you)
• Option B (editable install):
  • pip install -e '.[providers]'
  • Then run from anywhere.

.env example

OPENAI_API_KEY=sk-...

Support matrix (v1)

• OpenAI (GPT-4/5 and o-series): completions, typed commands, ask, streaming (no tools in stream), tool-calling, structured JSON for object schemas, tool-loop cap.
• Anthropic (Claude 3/3.5): completions and tool-calling loop (no streaming yet).
• Google (Gemini 1.5): basic completions (no tools/streaming in scaffold). Uses google-genai by default; google-generativeai supported via legacy extra.
• Ollama (local): basic completions via model="ollama:<name>" (no tools/streaming in scaffold).
• ReAct fallback: not implemented yet (planned for local models/LLMs without native tools).

Install options

• Base: pip install alloy-ai (includes OpenAI + python-dotenv).
• All providers: pip install 'alloy-ai[providers]' (OpenAI, Anthropic, Gemini via google-genai, Ollama).
• Specific extras: pip install 'alloy-ai[anthropic]', pip install 'alloy-ai[gemini]', pip install 'alloy-ai[ollama]'. Documentation
• Full docs: https://openai.github.io/alloy-py/

Releases

• Changelog: CHANGELOG.md
• Publishing: Create a tag like v0.1.1 on main — CI builds and uploads to PyPI (needs Trusted Publishing for alloy-ai or a configured token).