mittal-ai

Mittal Analytics' reusable AI harness. Install it as mittal-ai and import it as mittal_ai. It provides:

• streaming, non-streaming and structured LLM responses;
• tool-call handling, reusable tool helpers and message-history repair;
• token-cost calculation;
• OpenRouter routing and provider preferences;
• compatibility fixes for Chinese models; and
• dj-evals events for model requests and tool calls.

The application keeps its API keys. The model declares which provider and base URL the harness should use:

from mittal_ai import AIModel, get_client, get_structured_response
from pydantic import BaseModel


class Summary(BaseModel):
    text: str


model = AIModel(
    name="openai/gpt-5.4",
    api_key="...",
    provider="openrouter",
    base_url="https://openrouter.ai/api/v1",
    input_tokens_cost_usd=2.5,
    input_tokens_cached_cost_usd=0.25,
    output_tokens_cost_usd=15,
    output_tokens_reasoning_cost_usd=15,
)

async with get_client(model) as client:
    async for event in get_structured_response(
        client=client,
        ai_model=model,
        input=[{"role": "user", "content": "Summarise this."}],
        tools=[],
        text_format=Summary,
        reasoning_effort="low",
    ):
        print(event)

The main public entry points are get_response, get_streaming_response and get_structured_response.

Tool helpers

Use format_docstring to customize a reusable tool description and partial_with_doc to bind arguments that should not be exposed to the model:

from mittal_ai import format_docstring, get_streaming_response, partial_with_doc


@format_docstring(entity="company")
def answer_question(_company_id: int, question: str) -> str:
    """Answer a question about a {entity}.

    Args:
        - question: Question asked by the user.
    """
    return f"{_company_id}: {question}"


company_tool = partial_with_doc(answer_question, _company_id=123)

async for event in get_streaming_response(
    user="user-123",
    ai_model=model,
    messages=[{"role": "user", "content": "What changed?"}],
    tools=[company_tool],
    reasoning_effort="low",
):
    print(event)

company_tool preserves the tool metadata while exposing only question; the bound company ID remains private.

Publishing a new version

The release workflow publishes version tags to PyPI using trusted publishing. From a clean working tree, publish the next patch release with:

uv version --bump patch
version=$(uv version --short)

uv run pytest
uv run pre-commit run --all-files

git add pyproject.toml uv.lock
git commit -m "Release version $version"
git tag -a "v$version" -m "v$version"
git push origin main "v$version"

The tag must match the version in pyproject.toml, with a v prefix. Use --bump minor or --bump major when appropriate. PyPI versions are immutable, so every release needs a new version.