alloy-ai 0.1.4

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

Alloy (Python)

Python for logic. English for intelligence.

Alloy helps you integrate intelligence into deterministic programs with typed Python functions. Declare a function with @command(output=...), and Alloy routes to an LLM provider, enforces structured outputs, and returns typed results you can depend on.

License: MIT

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

Example: CSV to API

from alloy import command
import pandas as pd

@command(output=list[dict])
def csv_to_api(df: pd.DataFrame, endpoint_example: str) -> str:
    """Intelligently map CSV columns to API format."""
    return f"Map this data {df.head()} to match API: {endpoint_example}"

df = pd.read_csv("messy_customer_data.csv")
api_calls = csv_to_api(df, "POST /customers {fullName, emailAddress, subscriptionTier}")
for payload in api_calls:
    requests.post("https://api.your-saas.com/customers", json=payload)

This example maps rows from a DataFrame into request payloads for an API, returning a typed list[dict] you can post.

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 extract_price(text: str) -> str:
    """Extract price from text."""
    return f"Extract the price (number only) from: {text}"

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

Enforcing outputs

• Alloy biases models to return the expected shape and uses provider structured outputs where available. If parsing still fails, you get a clear error.
• Docs: https://lydakis.github.io/alloy/outputs/

Notes

• OpenAI backend is implemented for sync/async/streaming.
• Streaming with tools is not yet supported.
• Structured outputs: Alloy uses provider JSON Schema features (OpenAI/Anthropic/Gemini) and prompt guardrails. See Enforcing outputs above.
• 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-sonnet-4-20250514 (or another Claude like claude-3.7-sonnet). Run pytest -q — Anthropic integration tests auto-enable.
• Gemini: Set GOOGLE_API_KEY and ALLOY_IT_MODEL=gemini-2.5-pro (or gemini-2.5-flash). 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 3.7 / Sonnet 4 / Opus 4/4.1): completions and tool-calling loop (no streaming yet).
• Google (Gemini 2.5 Pro/Flash): basic completions (no tools/streaming in scaffold). Uses google-genai by default.
• 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://lydakis.github.io/alloy/

Why Alloy vs X

• Raw SDKs: Minimal glue, limited structure handling. Alloy provides typed outputs, provider routing, and a simple tool loop.
• LangChain: Rich orchestration features and chains. Alloy stays minimal: Python functions that return typed results without introducing a new framework.
• Instructor/Pydantic: Strong for OpenAI JSON typing. Alloy generalizes the idea across providers (OpenAI, Anthropic, Gemini), adds tools/retry/routing, and degrades gracefully when structure is not enforced.
• DSPy/Program synthesis: Optimizes pipelines and prompts. Alloy focuses on straightforward, production‑oriented building blocks: short prompts, typed outputs, and predictable defaults.
• Guidance/templating: Emphasizes prompt templates. Alloy emphasizes typed commands and provider structured outputs with clear error handling.
• Summary: Small API surface, provider‑agnostic backends, typed outputs by default, and optional tools — compose with normal Python.

LOC comparison (CSV → API payloads)

Raw SDK (conceptual):

import pandas as pd
from openai import OpenAI

client = OpenAI()
df = pd.read_csv("customers.csv")
messages = [{"role": "user", "content": f"Map data {df.head()} to {{fullName, emailAddress}}"}]
resp = client.chat.completions.create(model="gpt-4o", messages=messages)
payloads = json.loads(resp.choices[0].message.content)

Alloy:

from alloy import command
import pandas as pd

@command(output=list[dict])
def csv_to_api(df: pd.DataFrame, example: str) -> str:
    return f"Map data {df.head()} to {example}"

payloads = csv_to_api(pd.read_csv("customers.csv"), "POST /customers {fullName, emailAddress}")

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).