arxiv Tool

If you're building a research-assistant style agent in ComfyUI - something that helps summarize papers, track a topic, or answer "what's the latest work on X" - it needs a way to actually search the literature, because an LLM's training data has a cutoff and its memory of specific papers is unreliable. arxiv Tool is comfyui_LLM_party's answer: a straight wrapper around arXiv's own search, exposed as something your agent can call mid-conversation.

How it works

Add it to your agent's tool list, and the LLM gets a callable function for searching arXiv. When a question in the conversation calls for looking something up - "find recent papers on diffusion distillation," say - the model decides to invoke this tool, arXiv's search API runs the query, and results come back into the conversation for the LLM to read and summarize. No API key required: arXiv's search is free and open, which makes this one of the lowest-friction tools in the whole pack to set up.

The inputs and outputs that matter

• query - the search string. In practice you'll usually leave this at its default and let the LLM populate it at call time, since the point of wiring this in as a tool is that the model writes its own search queries based on what the user asked.
• is_enable - turn the tool off without deleting the node.
• One output: tool - wire this into the tool-list input on your LLM/agent node, exactly like every other tool node in the pack.

How to install it

• ComfyUI Manager - search "comfyui_LLM_party", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/heshengtao/comfyui_LLM_party, then restart.

Then run pip install -r requirements.txt from inside the comfyui_LLM_party folder, in your ComfyUI Python environment (python_embeded\python.exe -m pip install -r requirements.txt for portable Windows). This node needs no external account or key - arXiv's API is public - so beyond the pack's own baseline dependencies, there's nothing extra to configure to get it working.

Common issues & troubleshooting

The LLM never uses the tool. Function-calling models decide whether to call a tool from its description and the conversation context - if your system prompt or persona doesn't hint that the agent is a research assistant, the model may just answer from memory instead of searching. A line in your persona nodes along the lines of "search arXiv when asked about recent research" helps it reach for the tool more reliably.

Results feel generic or off-target. The tool is only as good as the query the LLM writes. If it's constructing vague queries, that's a prompting problem upstream - steer the agent (via the persona or system prompt) toward writing more specific search terms, not something you fix on this node.

Nothing happens at all. If arxiv_tool isn't even showing up in your node browser, that's the pack failing to load rather than this node specifically - a real, current issue where comfyui_LLM_party installs via Manager but individual nodes fail to import, usually from a CUDA/PyTorch mismatch with no clear compatibility list published. Check that first before assuming this node is the problem.