Add web search to Ollama with self-hosted Siftdog

In short: Run Ollama and Siftdog on your machine, then give a tool-capable Ollama model a small Python function that calls Siftdog’s /search endpoint. No third-party search or LLM API key is needed—but the web-search part is not offline.

What stays local—and what doesn’t

Ollama runs the model on your machine, and Siftdog runs the search API on infrastructure you control. The example below uses Siftdog’s REST API directly; it doesn’t need an MCP client or a hosted search service.

There is one important boundary: Siftdog uses SearXNG to query public search engines. Your search request needs internet access and is sent out to those engines. This is a local-model, self-hosted search setup—not a way to search the web privately or without an internet connection.

You don’t need an external search key or ANTHROPIC_API_KEY. The latter is only relevant to Siftdog’s optional include_answer feature; here, Ollama receives the search results and writes the response. By default, Siftdog authentication is disabled. You can set API_KEYS if you want to protect your server with a local bearer token.

Start Siftdog

Clone the Siftdog repository, then start the Docker Compose stack. It includes SearXNG for search:

git clone https://github.com/khsarvar/siftdog && cd siftdog
cp .env.example .env
docker compose up -d

The example .env doesn’t need third-party keys for this walkthrough. Siftdog’s API will be available at http://localhost:8000.

Prepare Ollama

Install Ollama for your operating system, then pull and start a model that supports tool calling. This example uses Qwen 3:

ollama pull qwen3
ollama run qwen3

Ollama’s tool-calling support lets a model request a function and use the result in a follow-up response. The function below connects that mechanism to Siftdog. See Ollama’s official tool-calling documentation for supported models and details.

Connect Ollama to Siftdog

Install the Ollama Python library and requests, then save this as search.py. The web_search function sends a query to your Siftdog server; Ollama can choose to call it when answering the prompt.

pip install ollama requests

import requests
from ollama import chat


def web_search(query: str) -> str:
    """Search the web through the local Siftdog server."""
    response = requests.post(
        "http://localhost:8000/search",
        json={"query": query, "search_depth": "advanced", "max_results": 5},
        timeout=60,
    )
    response.raise_for_status()
    return response.text


messages = [{
    "role": "user",
    "content": "Find the latest Python release and summarize it with source URLs.",
}]
response = chat(model="qwen3", messages=messages, tools=[web_search])
messages.append(response.message)

if response.message.tool_calls:
    for call in response.message.tool_calls:
        if call.function.name == "web_search":
            result = web_search(**call.function.arguments)
            messages.append({
                "role": "tool",
                "tool_name": call.function.name,
                "content": result,
            })
    response = chat(model="qwen3", messages=messages, tools=[web_search])

print(response.message.content)

If you set API_KEYS in Siftdog, add an Authorization: Bearer <key> header to the requests.post call. Keep that token on your own machine; it is for access to your Siftdog server, not a third-party search key.

Limits

Search depends on SearXNG and public search engines, which can rate-limit requests. If all configured engines are failing, Siftdog returns an error rather than an empty result that could be mistaken for “nothing found.” Ollama models can also vary in how reliably they decide to call tools, so choose a tool-capable model and check that the response includes sources.

If you want the full self-hosted API setup without the Ollama adapter, see how to self-host a Tavily alternative.