🛠️ Tool: Web Search 🛠️

A local LLM's knowledge stops the day its training data does, and no amount of prompting fixes that - ask a Qwen or Llama model what today's date is and it'll either refuse or confidently guess wrong. This node exists to fix exactly that gap: it gives an OllamaAgent the ability to actually query the live web mid-conversation and fold real results back into its answer, rather than making things up.

It doesn't do anything by itself - it defines a tool the agent can call. Whether it actually calls it depends on the model and the system prompt telling it to prefer tools over guessing (which is exactly what OllamaAgent's default system_context nudges it to do).

How it works

You get two search backends, chosen via search_provider:

• DuckDuckGo (free) - works out of the box, no signup, no key. This is the sane default for anyone just trying the feature.
• Ollama API - hits Ollama's own hosted search endpoint, which the pack's author describes as more accurate, but it requires a free API key from ollama.com/settings/keys and only makes sense once you're actually using Ollama's cloud models.

When the agent decides to call the tool, it gets back a handful of search-result snippets (however many max_results allows), which get fed back into the agent's context so it can reason over them before answering.

The inputs and outputs that matter

• tool_name (default search_internet) - this is literally the name the agent sees and reasons about internally. The node's own tooltip suggests renaming it to something like google_search or web_lookup if that helps the model understand when to reach for it - some models respond better to a more descriptive name than the generic default.
• search_provider - DuckDuckGo or Ollama API, as above. Start with DuckDuckGo; it's free and needs zero setup.
• max_results (default 5, capped at 10) - how many snippets come back. Fewer results means less context the agent has to chew through, which matters if you're running a smaller model with a tight context window.
• ollama_api_key (optional) - only used when search_provider is set to Ollama API; leave it blank for DuckDuckGo.

Output is a single tool value, type OLLAMA_TOOL - wire it directly into OllamaAgent's tools input, or into OllamaToolCombine if you're stacking this alongside other tools.

How to install it

Same pack, same setup as every node here. Via ComfyUI Manager, search "ComfyUI-Ollama-Describer" and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/alisson-anjos/ComfyUI-Ollama-Describer.git
pip install -r requirements.txt   # or install.bat on Windows

Restart ComfyUI. No extra dependencies specific to web search - DuckDuckGo needs no key at all, and the Ollama API path only needs the free key mentioned above.

Common issues & troubleshooting

The most common disappointment isn't an error, it's the agent just not using the tool - that's a model problem, not a node problem. Tool calling needs a model that actually supports it well (the README specifically calls out Llama 3.1 and Qwen 3.5 as tuned for tool use); a small or older model may ignore the tool entirely and hallucinate an answer instead. If you're getting no results at all rather than an ignored tool, check max_results isn't set unreasonably low and that DuckDuckGo isn't being rate-limited by your network - it's free, but free search endpoints do throttle heavy or repeated use. If you switched to the Ollama API provider, an empty or invalid key will simply fail that call - DuckDuckGo remains the fallback with zero config if you don't want to bother with a key at all.