> ## Documentation Index
> Fetch the complete documentation index at: https://docs.appliedaifoundation.org/llms.txt
> Use this file to discover all available pages before exploring further.

# PDF Vision Extractor

> Tiled AI-vision extraction for accurate Q&A on engineering diagrams, maintenance tables, and complex PDF pages.

## What it does

The **PDF Vision Extractor** is the precision counterpart to [PDF to Markdown](/skills/pdf-to-markdown). Where PDF-to-Markdown is built for bulk indexing, this skill is built for **accurate interactive extraction** from a small number of pages — engineering diagrams, maintenance schedules with dense tables, scanned forms with poor text quality.

It splits each page into overlapping tiles, runs vision inference on each tile, and merges results back into clean structured output. The tiling preserves detail that whole-page vision typically loses.

## When to use it

* "What does this diagram show?"
* "Read this table from the PDF"
* "Extract pages 5–10"
* "What are the maintenance intervals in this manual?"
* Standard text extraction returned garbled results

## When **not** to use it

For converting whole document libraries for search, use [PDF to Markdown](/skills/pdf-to-markdown). It's faster and produces the search index this skill does not.

## Comparison

|               | PDF Vision Extractor            | PDF to Markdown               |
| ------------- | ------------------------------- | ----------------------------- |
| Purpose       | Accurate Q\&A on specific pages | Bulk indexing for search      |
| Processing    | Tiled vision with overlap       | Whole-page vision             |
| Accuracy      | Highest — recovers fine detail  | Good — fast at scale          |
| Output        | One consolidated markdown       | Per-page files + search index |
| Query-focused | Yes                             | No                            |

## How it works

1. Renders each requested PDF page at high resolution.
2. Splits each page into overlapping tiles so detail at tile boundaries isn't lost.
3. Runs vision inference on each tile with a layout-aware prompt.
4. Merges tile output, deduplicating overlap, into clean per-page markdown.
5. Optionally answers a specific question using only the extracted content.

## Modes

* **Single page** — extract one specific page
* **Page range** — extract a contiguous range
* **Query mode** — extract and immediately answer a focused question against the extracted content

## Why tiling matters

Whole-page vision tends to summarise and miss small text, fine table cells, and dimensions on engineering drawings. Tiling forces the model to attend to local detail — at the cost of slower processing, which is acceptable for a few critical pages.

## Related skills

* **pdf-to-markdown** — bulk converter for full document libraries
* **search-indexed-documents** — search across the converted library