Top 200 ReposEXPLORER
Browse agents, categories and pages
Back to results
Planned, not doneThe source describes an intention, not a finished result.

Let a model search, read and cite web pages on its own in a self-hosted Open WebUI

In Open WebUI’s native tool-calling mode, a model decides when to search, opens the pages it wants to read in full and searches again, and only pages it actually opened are listed as citations.

Try the promptRead the original (opens in a new tab)
Upvote 0. Sign in to upvoteSign in to upvote · One vote per signed-in person. Site votes are separate from GitHub stars.

The problem

How it works

Open WebUI documents two built-in tools for web research. search_web returns titles, links and snippets from the search engine you configure, without storing anything in a vector database. fetch_url opens a page and puts its extracted text into the model’s context; PDFs at a link are read as documents. With native function calling, which the docs say is now the default, the model chains these steps itself: search broadly, notice gaps, read the most promising pages, and search again before answering. To set it up, an administrator configures a search engine, enables the Web Search capability and default feature on the model, and checks that function calling is set to Native.

Results and limitations

This is Open WebUI’s own documentation with a worked example but no measured results, so it is marked planned.

Before you try it

The docs say the model needs reliable tool calling and list both hosted models and local ones, such as Gemma 4 31B or Qwen 3.6 27B, that meet the bar; very small models struggle with the multi-step loop. Search snippets are no longer offered as citations, so ask the model to open a page before relying on it. Set a fetch length limit if long pages fill the context window. The quality of results depends heavily on the search engine you connect. Results depend on your setup and were not independently reproduced by Top 200 Repos.

Tools and repositories

Tools
Open WebUI · A search engine provider · A model with native tool calling
Prerequisites
An Open WebUI instance with admin access and a configured web search engine.

Try it yourself: Research a question and cite only what you read

Fill the fields to adapt the prompt. Empty fields stay as visible [placeholders].

Preview
Research this question: [Question]. Search the web, then open and read at least [Minimum pages to read] pages from different publishers before answering, preferring [Preferred sources]. Answer in [Answer format]. Cite only pages you opened, and say clearly where sources disagree or where you could not find an answer.

Open the cited pages yourself before you act on the answer.

Related videos

About open-webui/open-webui

About open-webui/open-webui

Independent creators. Listing a video implies no endorsement, sponsorship or partnership.

Sources

  1. Agentic Search & URL Fetching (opens in a new tab)

    Project documentation · docs.openwebui.com · Undated · Source checked 12 Nov 2026, 18:05 UTC

    What this source supports
    • In native mode the model decides if and when to search, using search_web for snippets and fetch_url to read full pages
    • search_web does no chunking or vector storage, and search hits are not listed as citations; only pages opened with fetch_url are citable
    • Native function calling is the default; Legacy mode is the unsupported opt-out
    • Setup: configure a search engine, enable the Web Search capability and default feature on the model, and confirm Function Calling is Native
    • Lists hosted and local models such as Gemma 4 31B and Qwen 3.6 27B that meet the minimum bar, and a fetch content length limit
Report a problem

Reported by source

Extract invoice fields locally with a vision model on Ollama

Stephen Sebastian replaced a paid cloud OCR API with Gemma 4 E4B on Ollama to pull date, amount and description from contractor invoice photos on his laptop.

dev.toollama/ollama
Reported by source

Get a new OCR model running on unfamiliar GPU hardware with Claude Code in a sandbox

Simon Willison gave Claude Code full permissions inside a Docker container on an NVIDIA Spark and had it get DeepSeek-OCR working; it took four prompts and under 40 minutes, with one nudge from him.

simonwillison.net · 20 Oct 2025anthropics/claude-code, pytorch/pytorch, huggingface/transformers
Planned, not done

Give your agent a local model that can call tools, served by llama.cpp’s llama-server

llama-server from llama.cpp speaks the OpenAI chat API, and with its --jinja flag it turns model output into OpenAI-style tool calls, so agents and libraries built for hosted APIs can run on a model on your own machine.

github.comggml-org/llama.cpp
Reported by source

Run an open-weight model locally and call it from your code

Install Ollama, pull an open model and call it through the local REST API so prompts stay on your machine.

github.comollama/ollama
Planned, not done

Run OpenAI’s Codex CLI on open models served by Ollama

Ollama’s docs show how to point Codex CLI at local tool-capable models such as gpt-oss with one launch command or a saved profile, switch models inside a session and turn off web search.

docs.ollama.comopenai/codex, ollama/ollama
Reported by source

Answer technical questions by letting async agents run experiments

Simon Willison sends research questions to cloud coding agents such as Claude Code for web and Codex Cloud, which write and run experiments in a dedicated repo and open a PR.

simonwillison.net · 6 Nov 2025anthropics/claude-code, openai/codex