llm-markdown 0.3.9

Turn Python functions into typed LLM calls using docstrings as prompts

llm-markdown

LLM calls as Python functions. Write a docstring, add a type hint, done.

import os

from llm_markdown import prompt
from llm_markdown.providers import OpenAIProvider

provider = OpenAIProvider(
    api_key=os.environ["OPENAI_API_KEY"],
    model="gpt-4o-mini",
)

@prompt(provider)
def summarize(text: str) -> str:
    """Summarize this text in 2 sentences: {text}"""

result = summarize("Long article text here...")
# "The article discusses... In conclusion, ..."

How it works

Three rules:

1. The docstring is the prompt. Use {param} to interpolate function arguments.
2. The return type controls the output format. -> str gives plain text. -> MyModel gives validated structured output.
3. Image parameters are attached as vision inputs automatically -- they don't appear in the docstring.

That's it. No configuration flags, no prompt templates, no output parsers. The function signature is the configuration.

Installation

pip install llm-markdown[openai]

For development with full provider + test support:

python -m venv .venv
source .venv/bin/activate
pip install -e ".[test,all]"

Provider extras: openai, anthropic, gemini, openrouter.

Other extras: langfuse (observability), all (all providers + langfuse), test (pytest suite).

Agent streaming (tools + reasoning)

For Hof-style tool loops, use stream_agent_turn with ReasoningConfig so OpenAI and Anthropic apply the same native / off policy. See docs/agent-streaming.md for when AgentReasoningDelta appears and provider capability notes.

Provider support

Provider	Included	Native structured output	Images	Streaming	Extra
OpenAI	Yes	Yes (response_format)	Yes	Yes	openai
Anthropic	Yes	Yes (tool schema)	Yes (data URI images)	Yes	anthropic
Google Gemini	Yes	Yes (response_schema)	Yes (data URI images)	Yes	gemini
OpenRouter	Yes	Yes (OpenAI-compatible schema)	Model-dependent	Yes	openrouter

Install one provider:

pip install llm-markdown[anthropic]
pip install llm-markdown[gemini]
pip install llm-markdown[openrouter]

Or install all providers:

pip install llm-markdown[all]

Instantiate providers:

import os

from llm_markdown.providers import (
    OpenAIProvider,
    AnthropicProvider,
    GeminiProvider,
    OpenRouterProvider,
)

openai_provider = OpenAIProvider(api_key=os.environ["OPENAI_API_KEY"], model="gpt-4o-mini")
anthropic_provider = AnthropicProvider(api_key=os.environ["ANTHROPIC_API_KEY"], model="claude-3-5-sonnet-latest")
gemini_provider = GeminiProvider(api_key=os.environ["GOOGLE_API_KEY"], model="gemini-2.0-flash")
openrouter_provider = OpenRouterProvider(
    api_key=os.environ["OPENROUTER_API_KEY"],
    model="inception/mercury-2",
    app_name="llm-markdown",
    app_url="https://example.com",
)

OpenRouter access is configured in three places:

• .env: OPENROUTER_API_KEY
• Provider instance: OpenRouterProvider(model="...")
• Integration test model list: OPENROUTER_TEST_MODELS in tests/test_integration_multi_providers.py

Structured output

Return a Pydantic model and the response is validated automatically:

from pydantic import BaseModel

class ReviewAnalysis(BaseModel):
    sentiment: str
    rating: float
    key_points: list[str]

@prompt(provider)
def analyze_review(text: str) -> ReviewAnalysis:
    """Analyze this movie review:
    - Overall sentiment (positive/negative/neutral)
    - Rating on a scale of 1.0 to 5.0
    - Key points

    Review: {text}"""

result = analyze_review("A groundbreaking sci-fi film...")
result.sentiment    # "positive"
result.rating       # 4.5
result.key_points   # ["groundbreaking visual effects", ...]

The library generates a JSON schema from the Pydantic model and uses the provider's native structured output (e.g. OpenAI's response_format). If the provider doesn't support it, it falls back to JSON prompting automatically.

List[...] and Dict[...] work the same way:

from typing import List

@prompt(provider)
def list_steps(task: str) -> List[str]:
    """List the steps to complete this task: {task}"""

list_steps("bake a cake")
# ["Preheat oven to 350F", "Mix dry ingredients", ...]

Images

Image parameters are detected by type and attached to the API call as vision inputs. The docstring is the text part of the prompt:

from llm_markdown import prompt, Image

@prompt(provider)
def answer_about_image(image: Image, question: str) -> str:
    """Answer this question about the image: {question}"""

answer_about_image(
    image=Image("https://example.com/chart.png"),
    question="What trend does this chart show?",
)

Image accepts URLs, local file paths, base64 strings, or data URIs. Non-image content types and payloads above 20MB are rejected. Use List[Image] for multiple images.

Image generation

Use the high-level helper for provider-backed image generation:

from llm_markdown import generate_image

image_result = generate_image(
    provider=openrouter_provider,
    prompt="A minimalist watercolor mountain scene",
    model="openai/gpt-image-1",
)
print(image_result["images"][0]["url"])

Async variant:

from llm_markdown import generate_image_async

image_result = await generate_image_async(
    provider=openrouter_provider,
    prompt="A retro sci-fi city skyline",
    model="openai/gpt-image-1",
)

Streaming

@prompt(provider, stream=True)
def tell_story(topic: str) -> str:
    """Tell a short story about {topic}."""

for chunk in tell_story("a robot learning to paint"):
    print(chunk, end="", flush=True)

Structured event streaming is also available:

from pydantic import BaseModel

class Answer(BaseModel):
    value: str

@prompt(provider, stream=True, stream_mode="json_events")
def stream_answer(question: str) -> Answer:
    """Return a JSON answer for: {question}"""

for event in stream_answer("What is 2+2?"):
    print(event["type"])

Async

Async functions work the same way:

@prompt(provider)
async def analyze(text: str) -> str:
    """Analyze: {text}"""

result = await analyze("some text")

Sessions (multi-turn)

from llm_markdown import Session

session = Session(provider, max_messages=12)

@session.prompt()
def ask(question: str) -> str:
    """Answer this: {question}"""

ask("What is retrieval-augmented generation?")
ask("Now explain it to a beginner.")

Session keeps shared chat history and supports optional max_messages/max_tokens trimming.

Generation controls and metadata

Pass default generation controls at decoration time and override per call with _llm_options:

from llm_markdown import prompt, PromptResult

@prompt(
    provider,
    generation_options={"temperature": 0.4, "max_tokens": 300},
    return_metadata=True,
)
def summarize(text: str) -> str:
    """Summarize: {text}"""

result: PromptResult[str] = summarize(
    "Some article",
    _llm_options={"temperature": 0.2},
)
print(result.output)
print(result.metadata)  # provider/model/response_id/usage

Observability with Langfuse

Wrap any provider with LangfuseWrapper to log every call:

import os

from llm_markdown.providers import OpenAIProvider, LangfuseWrapper

provider = LangfuseWrapper(
    provider=OpenAIProvider(api_key=os.environ["OPENAI_API_KEY"]),
    secret_key=os.environ["LANGFUSE_SECRET_KEY"],
    public_key=os.environ["LANGFUSE_PUBLIC_KEY"],
    host="https://cloud.langfuse.com",
)

@prompt(
    provider,
    langfuse_metadata={"category": "reviews", "use_case": "sentiment"},
)
def analyze(text: str) -> str:
    """Analyze: {text}"""

Custom providers

Subclass LLMProvider to use any LLM backend:

from llm_markdown.providers import LLMProvider

class MyProvider(LLMProvider):
    def complete(self, messages, **kwargs):
        ...  # return response string

    async def complete_async(self, messages, **kwargs):
        ...  # return response string

    # Optional -- enables native structured output.
    # Without this, the decorator falls back to JSON prompting.
    def complete_structured(self, messages, schema, **kwargs):
        ...  # return parsed dict

Built-in providers:

• OpenAIProvider: OpenAI models (GPT-4o, GPT-5, o1/o3/o4) with automatic token parameter detection.
• AnthropicProvider: Claude models with native structured output via tool schema.
• GeminiProvider: Gemini models with native structured output via response schema.
• OpenRouterProvider: OpenAI-compatible models routed through OpenRouter.
• RouterProvider: route/fallback wrapper across multiple providers.

Testing

python -m llm_markdown.preflight --strict
pytest -m "not integration"      # required quality gate before commit
cp .env.example .env           # fill provider keys
set -a; source .env; set +a
pytest -m integration           # optional real provider API tests
python -m build                 # required quality gate before push

If you installed with .[test,all], integration tests can also auto-read .env via python-dotenv.

Required keys for provider integration tests:

• OPENAI_API_KEY
• ANTHROPIC_API_KEY
• GOOGLE_API_KEY
• OPENROUTER_API_KEY

Integration tests run against a curated model set per provider and skip individual model cases if a model is not enabled for the key/account.

Troubleshooting

• ImportError for provider packages: install matching extras (llm-markdown[openai], llm-markdown[anthropic], llm-markdown[gemini], llm-markdown[openrouter]).
• Missing function docstring: every @prompt function needs a docstring prompt template.
• Structured outputs not supported by provider: the decorator automatically falls back to JSON prompting.
• stream=True returns a stream iterator and bypasses structured parsing.
• Image issues: URLs/local files must resolve to an image MIME type, and payloads above 20MB are rejected.

Security notes

• Keep secrets in environment variables (.env) and never hardcode keys in source.
• Do not commit .env files or raw credentials.
• Treat remote image URLs as untrusted input; prefer trusted sources for production.
• Restrict remote image hosts with LLM_MARKDOWN_IMAGE_URL_ALLOWLIST (comma-separated domains).
• Keep private-network URL blocking enabled (LLM_MARKDOWN_IMAGE_BLOCK_PRIVATE_NETWORKS=true).

Versioning and release

• Project version currently lives in llm_markdown/__init__.py and setup.py.
• Document user-visible changes in release notes or PR descriptions.
• Use docs/release-notes-template.md for release notes and breaking-change checklist.
• Before publishing, run pytest -m "not integration" and python -m build.

Additional docs

• docs/getting-started.md
• docs/providers.md
• docs/structured-output.md
• docs/images-and-multimodal.md
• docs/streaming-and-async.md
• docs/troubleshooting.md
• docs/security.md
• docs/contributing.md
• docs/versioning.md
• docs/operations-runbook.md
• docs/release-notes-template.md