llm-deepseek-websearch

An llm plugin that gives models two web tools, built on the DeepSeek Search and Fetch APIs:

• deepseek_web_search: search the web and return ranked results with summaries.
• deepseek_web_fetch: fetch one page and return its content as Markdown.

The tools work with any model that supports tools in llm, not only DeepSeek models.

We use this plugin for research as part of our article writing process for Kafkai Blog, Kafkai AI Media and PuchiDen Blog.

Installation

Install this plugin in the same environment as LLM. You need LLM 0.33 or newer and Python 3.10 or newer.

llm install llm-deepseek-websearch

To install the latest code from GitHub before a release reaches PyPI:

llm install https://github.com/kafkaigiken/llm-deepseek-websearch/archive/refs/heads/main.zip

API key

Create a key on the DeepSeek platform, then store it:

llm keys set deepseek     # paste the key

If no stored key exists, the plugin uses the DEEPSEEK_API_KEY environment variable. A stored key takes priority over the environment variable, so remove it with llm keys remove deepseek before relying on the variable. The deepseek key alias is the same one that the llm-deepseek model plugin uses, so one stored key serves both plugins.

Usage

Search:

llm -T deepseek_web_search "What changed in the latest Python release?" --td

Search, then read the best pages:

llm -T deepseek_web_search -T deepseek_web_fetch \
  "Find the DeepSeek API rate limits and quote the official page." --td

Fetch a page you already know:

llm -T deepseek_web_fetch "Summarize https://llm.datasette.io/en/stable/tools.html"

--td (--tools-debug) prints every tool call and its result, so you can see what the model searched for.

deepseek_web_search

Each result shows the title, URL, site name, date, relevance score (0 to 3, 1.5 or more is a good match), authority tier (0 to 4, 3 or more is highly authoritative) and a summary of about 600 to 800 characters.

The query supports these operators, and you can combine them:

The tool docstring lists the operators, so models use them without extra prompting.

deepseek_web_fetch

It returns the page title, the URL and the page content as Markdown. The content is not cut.

Errors and retries

• The plugin makes up to 3 attempts. It retries HTTP 408, 429 and 5xx responses, and also network failures. Other HTTP errors fail at once.
• Between attempts it waits for the Retry-After header, given as seconds or as an HTTP date, and never longer than 10 seconds. Without a usable header it backs off from 1 second.
• The model sees the HTTP status and DeepSeek's error message, for example DeepSeek Fetch API error 404: ... when a page has no content that DeepSeek can extract. The plugin cuts a long error body to 1,000 characters.
• A redirect fails with the new location in the message. The plugin never follows a redirect, so an API that moves a path is reported rather than silently used.
• A bad url fails before any request is sent, with a message naming the problem, for example url must start with http:// or https://.
• A malformed 200 response fails with a message saying the API may have changed, rather than looking like an empty search or an empty page.
• HTTP 402 means the account balance is too low. Add funds on the DeepSeek platform.

Pricing and limits

From the DeepSeek documentation on 2026-10-19:

• Search and Fetch each cost USD 0.75 per 1,000 successful requests. Errors, empty results and timeouts are free. count does not change the price.
• Each API allows 5 requests per second per account. Fetch also has a limit per domain.

Development

To set up this plugin locally, check out the code, then create a virtual environment:

cd llm-deepseek-websearch
python -m venv venv
source venv/bin/activate

Install the dependencies and the test dependencies:

llm install -e '.[test]'

Run the tests:

python -m pytest

The tests use a mock HTTP transport, so they never call the real API and need no key.

Releasing

The Publish Python Package workflow publishes to PyPI with trusted publishing, so the repository stores no PyPI token.

To release a new version:

1. Change version in pyproject.toml, then commit and push.
2. On GitHub, create a release with a new tag, for example 0.2.
3. The workflow runs the tests on Python 3.10 to 3.13, builds the package and uploads it to PyPI.

To check the package before a release, build it in a separate virtual environment:

python -m venv /tmp/buildenv
/tmp/buildenv/bin/pip install build twine
/tmp/buildenv/bin/python -m build --outdir /tmp/dist
/tmp/buildenv/bin/twine check --strict /tmp/dist/*