Checking the internet for information.

[fimileurgamer-design]

I’m really new to local llm’s and ai in general and I’d like to know if there is a way for me too get my local models to search the web on websites specified in my prompt and have the data be live and relevant for example I can ask it to search eBay for listings on new laptops.

additionally I’d like to know if there’s something similar to skills on Claude?

Any information is appreciated.

[rofl0r]

it's the agent harness that does all of the above things: a program that connects to your inference server (be it colibri, llama.cpp, or whatever) and runs the LLM in a loop when you prompt it. it also offers the tools the model can call, including potentially a web search tool. but with all the hostility in the web against bots it's difficult to find a search engine that let's you use it for free e.g. via curl, unless you host your own searx instance, or you pay for a dedicated API key for a special search engine service. some harnesses such as hermes install hundreds of megabytes of headless browser toolkits for its web tools, so to a website it looks like the agent is using a "legit", fingerprintable browser. this solution should mostly work, but it's quite resource heavy. for accessing a bloated, js-heavy site like ebay a headless browser is your only option anyway though. note that you most likely need a model with visual capabilities (llama.cpp based models usually offer a "mmproj" file for the visual cortex, if the model supports it), so it can recognize which buttons to click. i can imagine that tokenizing a visual representation of a website is also quite token heavy, so a model that runs at only 1 t/s will likely take a very long time do do anything on a headless browser.

[fimileurgamer-design]

Thanks so much I’ll research the topics that you mentioned.🫡