How do I make search with local models feel like ChatGPT / Codex?

I’m using Crawl4AI + SearXNG but it just feels so lackluster I don’t know why. Feels like the results the local model gets back for research are missing a lot and / or are older less relevant results. I’ll usually instruct saying something like “using the most up-to-date and relevant information across related websites, subreddits, forums, and etc find me the best way to _______” or something similar. With ChatGPT or Codex it basically always finds everything it needs to and gives me great instructions or information or research or whatever I’m looking for at the time. With a local model (like GLM 5.3 Flash or DeepSeek V4.1 Flash 1017) the answer doesn’t come back nearly as good, often missing obvious results or mixing in outdated information into the instructions or etc. Harnesses I’ve been using with local models are ZCode, DeepSeek Harness, and OpenCode. Is there a better (more complete and on par with ChatGPT) searching / research approach I’m missing? Is there some tuning of Crawl4AI + SearXNG to make it more how I’m expecting or potentially something I should add to this combination? Any input or help is appreciated. Thank you for your time it is appreciated!