
## What is Extract in Maxun?

**Extract** is Maxun’s core.

It emulates real user behavior in a real browser to collect structured data from any website — no code required.  
If a human can navigate a site, Extract can automate it.

This includes
1. JavaScript-heavy websites
2. Pagination and dynamic content
3. Real-world user flows

The output is clean, structured data you can use in APIs, spreadsheets, or downstream systems.

<br />
<hr />

## How Extract Works

Until now, Extract worked through **Recorder Mode**.

You would
1. Open a website
2. Click, scroll, and navigate as a real user
3. Capture the data you want
4. Save the flow as a reusable extraction robot

Maxun converts these actions into a **deterministic workflow** that can be replayed reliably at scale.

Recorder Mode is powerful — and it’s still here.

<br />
<hr />

## From Recording to Reasoning

This is where Extract evolves.

Instead of teaching the system *how* to extract data step-by-step, you can now describe *what* you want — and let the system reason about the rest.

We call this **AI Mode**.

<br />
<hr />

## Recorder Mode vs AI Mode

Both modes solve different problems.
<br />
### Recorder Mode
1. Fully deterministic
2. Best for complex flows and long-running jobs
3. Ideal when reliability matters most
<br />
### AI Mode
1. Intent-driven
2. Fastest way to get started
3. Best for exploratory extraction
<br />
They are not replacements for each other.  
They run on the same browser-based engine.  
You choose based on the problem.

<br />
<hr />

## AI Mode: Describe What You Want

With **AI Mode**, you describe the data you want in natural language.

The LLM reasons about
1. Page structure
2. Navigation
3. Field boundaries
4. Extraction logic

The result is clean, structured output — just like Recorder Mode.

<br />
<hr />

## Support for Closed and Open-Source LLMs

AI Mode supports multiple model backends.

1. **Local LLMs via Ollama** for privacy and self-hosting
2. **Cloud models** (OpenAI, Anthropic) for higher accuracy on complex pages

You control where inference runs and which model is used.

<br />
<hr />

## Example: Extract Top 50 Movies from IMDb

Prompt:
Extract the names, ratings, and duration of the top 50 movies.

<video
  src="/maxun_ai_mode_extraction.mp4"
  controls
  autoPlay
  loop
  muted
  style={{ borderRadius: "8px", width: "100%", margin: "1.5rem 0" }}
/>
<br />
*AI Mode extracting the top 50 movies from IMDb using a natural language prompt.*
<br />

Without recording any steps, Extract:
1. Navigates the page
2. Identifies the correct list
3. Extracts structured fields
4. Returns clean data

<br />
<hr />

## Where This Is Going

Recording gives determinism. AI adds reasoning.

Together, they make Extract flexible enough for exploration and reliable enough for production.

This is the next step in how we think about web extraction — and it’s just getting started.


From Recording to Reasoning with AI


The status quo of web scraping is broken for AI. For a decade, web extraction was a war over CSS selectors and DOM structures. We wrote brittle scrapers that broke the moment a `div` turned into a `span`. 

But as we enter 2026, the bottleneck isn't getting the data : it’s the **quality and density** of that data when it hits an LLM’s context window.

If you are still feeding raw HTML to your RAG pipelines or AI agents, you are paying a "tax" in tokens, latency, and hallucinations.

Here is why **Markdown** is the missing link in the modern AI data stack.

---

## The Token Tax: HTML is 90% Noise

Large Language Models don't read web pages; they process tokens. A standard e-commerce product page can easily reach **150KB of HTML**, which translates to roughly **40,000+ tokens**.

When you convert that same page to clean, semantic Markdown:

1. **Size drops by 95%:** You go from 40,000 tokens to ~2,000.
2. **Cost Efficiency:** You can process 20x more pages for the same API cost.
3. **Signal-to-Noise Ratio (SNR):** You strip away `<script>`, `<style>`, and nested `<div>` that forces the model's attention mechanism to work harder for less signal.

<table> <thead> <tr> <th align="left">Data Format</th> <th align="center">Avg. Tokens per Page</th> <th align="center">Estimated Cost (GPT-4o)</th> <th align="center">Cost Efficiency</th> </tr> </thead> <tbody> <tr> <td><strong>Raw HTML</strong></td> <td align="center">45,000</td> <td align="center">$0.1125</td> <td align="center">Baseline</td> </tr> <tr> <td><strong>Clean Markdown</strong></td> <td align="center">1,800</td> <td align="center"><strong>$0.0045</strong></td> <td align="center"><strong>96% Reduction</strong></td> </tr> </tbody> </table>

Note: These estimates are based on current 2026 pricing for GPT-4o at $2.50 per 1M input tokens. By distilling HTML into Markdown, you're effectively increasing your context window by 25x for the same price.*

---

## Structural Bias: LLMs are Native Markdown Speakers

LLMs are trained on the internet, which means they are trained on **GitHub, StackOverflow, and Technical Documentation**. The "Common Crawl" of high-quality reasoning data is written in Markdown.

Markdown provides **semantic hierarchy** that HTML obscures:

1. **Headers (#, ##):** Explicitly define the "parent-child" relationship of ideas.
2. **Tables (|):** Allow models to perform "columnar reasoning" (e.g., comparing prices across rows) without getting lost in `<tr>` and `<td>` nesting.
3. **Bullet Points (-):** Signal distinct entities or steps in a process.

When a model sees a Markdown header, it understands it as a **context anchor**. In raw HTML, that same header is just another node in a 50-level deep DOM tree.

---

## RAG Accuracy: The "Chunking" Problem

Most RAG pipelines use "Naive Chunking"- splitting text every 500 characters.

1. **The HTML Failure:** A split might happen in the middle of a `<table>` tag, effectively destroying the data's meaning for the vector database.
2. **The Markdown Solution:** Markdown allows for **Semantic Chunking**. You can split data at the `#` or `##` boundaries. This ensures that every chunk in your vector store is a coherent, self-contained unit of information.

Technical Insight: "Header-Aware Chunking" in Markdown-based RAG pipelines has been shown to improve retrieval accuracy by **40% to 60%** because the embeddings capture the *contextual intent* of the section rather than just random word proximity.*

---
## The Path Forward: Data is the New Code

We are moving toward a future where the "Browser" is just an OS for AI Agents.

The goal of data extraction in 2026 isn't just to "have" the data - it's to make it usable for the machines that will process it. High-density, structured Markdown is the only way to make LLMs smarter, faster, and cheaper to run.

We are building the future of AI-native extraction to bridge the gap between the messy web and the clean context windows your models deserve.

Ready to turn the web into your personal database? 
[Get Started For Free!](https://www.google.com/search?q=https://app.maxun.dev/register)

---

## Join the Community

We are building the future of no-code, AI-native extraction.

1. [Maxun GitHub Repository](https://www.google.com/search?q=https://github.com/getmaxun/maxun)
2. [Maxun Cloud](https://www.google.com/search?q=https://app.maxun.dev/register)


Blog

From Recording to Reasoning with AI

Why Markdown is the Secret to Better AI