What’s the cheapest way to give a local Llama 3 internet access? (SearXNG isn’t cutting it)
Posted by Old-Tumbleweed1422@reddit | LocalLLaMA | View on Reddit | 20 comments
Finally got Llama 3 70B running locally and wired up function calling so it can search the web. First tried self-hosting SearXNG, but the results are pretty messy. Then I tested Brave Search API, but the snippets are too short - the model just doesn’t get enough context to generate decent answers.
Looking for a cheap (ideally free for a side project) API that can quickly return useful chunks of website content instead of tiny snippets
What are you guys using?
Acrobatic_Stress1388@reddit
This guy is so 6 months ago
spambait-aspaaaragus@reddit
Help me out I don’t get it
my_name_isnt_clever@reddit
I'm guessing it's OP using Llama 3 70b. That's a grandpa model these days, especially since it was before mainstream tool calling so it's not very useful these days.
Blindax@reddit
In OpenWebUI: use a recent model in native tool calling mode. SearXNG works well once properly configured. Either way, configure a web loader so the model gets full page content instead of just snippets, that’s likely your real issue. If you skip function calling, also set up embedding and retrieval for the RAG pipeline.
jojotdfb@reddit
Llama 3 is a very old model. This space moves so fast. Think of each month as a year. Llama 3 might as well be 20 year old. If you can run Llama 3 70b, you could run Qwen 3.6 27b easily and get better results.
Risen_from_ash@reddit
You’re still using GPU based models?
dreamtheater2003@reddit
Using ddgs for websearch and trafilatura for fetch. Also integrated that successfully with openwebui for local llms with websearch. Works quite well, but also still looking at optimizing my setup further.
McZootyFace@reddit
Step one uninstall Llama 3.
mister2d@reddit
Searxng just returns search results and some metadata; like what you see from a web search. That's all you get to work with. It does not fetch web content.
graypasser@reddit
Just ask models to fetch actual website content?
my_name_isnt_clever@reddit
Have you actually tried that for research? Sure you can fetch docs with just
curlbut most sites with actual content block anything that simple. I was using jina.ai free for awhile, now I pay for Exa content retrieval only.UnWiseSageVibe@reddit
What was wrong with searing? It works fine with me
I use it together with firecrawl self hosted.
my_name_isnt_clever@reddit
I use SearXNG for free search results that I can customize in detail using its's URL filtering configuration. Then I use Exa pay as you go only for retrieval, as that's the hard part that every website is trying to prevent right now. It's been a couple weeks and I just checked my Exa balance, I've only spent $0.20 on it so far and every other part of my stack is FOSS self hosted.
starkruzr@reddit
... why
ubrtnk@reddit
I use a combination mcp server that I built in n8n that leverages searxng for the basic search function and then a 2nd pass with Jina.ai to read the url that was found on searxng.
Scared-Tip7914@reddit
Aight imma shamelessly plug my stuff here but if you want to search the web for free and get results and web search thats are actually relevant, not 69k tokens of bullcrap try this: https://github.com/MarcellM01/TinySearch. I made it so that no matter the question it keeps the response under 8k.
wotoan@reddit
Tavily works great for Qwen 3.6 and has a free tier. I’ve never come close to using it all up but I’m not a heavy user. Worth a shot though.
blackhawk00001@reddit
Open-websearch mcp works good with Claude cli, much faster than the built in search for local hosted backend
guigouz@reddit
I'm using https://exa.ai/ it works fine and has a free tier.
Doogie707@reddit
"Cheapest" nga are you PAYING FOR SEARCH IN THE BIG 2026??? LMAOOO😭
Okay here's a couple of options:
1 - Brave search /answers api - you can have your model just ask the ai instead of even searching, brave handles the search, ranking etc and gives your model the answers OR the links for it to explore.
However, this is how a boomer might do it, for you my zonked zillenial z-compadre, the answer is actually much simpler:
2 - Download EITHER (You CAN use both but thats bloated) browser-harness OR agent-browser. They both come with skills that the agent can invoke at runtime so you dont have to tell it anything except what its searching for.
Enjoy, and save your damn money lol