Context, Context Everywhere, Not a Drop to Use

This talk explores a simple problem statement: How can we efficiently find the right context in a sea of information (the internet) to answer a question we might have?

It goes through 5 approaches, each more refined than the last:

1. Ask an AI agent and rely on its default web search.

2. Stop guessing and fetch each page with cURL.

3. Render pages in a real browser and deep crawl the website using Crawl4AI.

4. Use an LLM to interpret each page’s clean markdown to find the right context

5. Let a small decision model (Jev) check each page first. The LLM then
   reads only the page that passes.

Final Result: One Correct Grounded answer for $0.00067 (2,050x fewer tokens than giving raw HTML to the LLM)

• How to cut tokens while web crawling.

• How to crawl JavaScript-heavy sites with Crawl4AI.

• How to cut LLM costs with a model cascade.

• How to deep crawl websites using a Best First Search Strategy.