websum
Summarise a web page or a YouTube video, with whichever model you want behind it.

PyPI · changelog · contributing

Summarising a page is a few lines of LangChain right up until you want the same call to also take a YouTube link, chunk a transcript that will not fit in a context window, run against a local Ollama on a laptop with no network, and then move to OpenAI without any of the calling code changing. That is the part this packages.

One object, Summarizer, takes a URL and works out whether it is a page or a video. The model behind it is a constructor argument, so swapping Ollama for OpenAI, or for something you wrote yourself, is a single line.

from websum import Summarizer, OllamaBackend

s = Summarizer(backend=OllamaBackend(model="llama3:instruct"))
print(s.summarize("https://www.youtube.com/watch?v=4pOpQwiUVXc"))

• Pages and YouTube through the same call. summarize(url) detects which it has. summarize_web and summarize_youtube are there when you already know.
• The backend is a Protocol, not a base class to inherit. Anything with a .build() returning a LangChain chat model satisfies it.
• Local by default. Ollama is a first-class backend, so nothing has to leave the machine, and no key is required to try it.
• Long inputs are chunked, so a two-hour transcript does not have to fit anywhere in one piece.
• A CLI, a library and a Gradio UI over the same code, and a py.typed marker so the API type-checks downstream.

pip install 'websum[ollama]'     # library + CLI, local models
pip install 'websum[openai]'     # OpenAI backend
pip install 'websum[ui,ollama]'  # with the Gradio web UI
pip install 'websum[all]'        # everything

Using uv:

uv add 'websum[ollama]'

Python 3.10 to 3.13. The base install carries no model client at all, which is why the backend is an extra: installing websum alone should not drag in an SDK you are not going to call.

As a library

from websum import Summarizer, OllamaBackend

s = Summarizer(backend=OllamaBackend(model="llama3:instruct"))
print(s.summarize("https://cobanov.dev/haftalik-bulten/hafta-13"))
print(s.summarize("https://www.youtube.com/watch?v=4pOpQwiUVXc"))
print(s.translate("Hello world", target_language="Turkish"))

Swap the backend without touching anything else:

from websum import Summarizer, OpenAIBackend

s = Summarizer(backend=OpenAIBackend(model="gpt-4o-mini"))

From the shell

websum summarize https://example.com
websum summarize https://example.com --backend openai --model gpt-4o-mini
websum translate "Hello world" --target-language Turkish
websum ui --port 7860

websum --help has the full reference.

In a browser

websum ui --port 7860

Every public name is re-exported from the top-level websum package and listed in __all__.

LLMBackend is a Protocol, so there is nothing to subclass and nothing to register. A class with a build() method already satisfies it:

from dataclasses import dataclass
from websum import LLMBackend, Summarizer

@dataclass
class MyBackend:
    def build(self):
        from langchain_anthropic import ChatAnthropic
        return ChatAnthropic(model="claude-sonnet-5")

assert isinstance(MyBackend(), LLMBackend)  # runtime Protocol check
s = Summarizer(backend=MyBackend())

The import lives inside build() on purpose: a backend that is never constructed should not cost an import of an SDK that may not be installed.

docker build -t websum .
docker run -p 7860:7860 websum

# when ollama is running on the host
docker run --network host -p 7860:7860 websum

The image starts websum ui by default.

git clone https://github.com/cobanov/websum
cd websum
uv sync --all-extras
uv run pre-commit install
uv run pytest
uv run ruff check .
uv run mypy src/websum

CI runs the 19 tests on 3.10, 3.11, 3.12 and 3.13, with ruff, ruff format --check and mypy on top. pytest-randomly shuffles the order every run, so a test that only passes because another one ran first fails here rather than later. CONTRIBUTING.md has the full guide.

The 0.1.x scripts under app/ are gone. Everything moved into the websum package behind a typed, importable API.

MIT. See LICENSE.