How to prepare a PDF for RAG without losing context

You have a 90-page operations manual and need to feed it into a RAG pipeline. The naive approach is to dump everything in and hope retrieval figures it out. What actually happens: chunks get ripped mid-sentence, tables lose their headers, and the retrieval step returns fragments that confuse the model more than they help.

Preparing a PDF for RAG is not about raw conversion. It is about preserving the document structure that gives text meaning.

Why raw extraction breaks context

PDF is a visual format. The internal structure stores text as positioned fragments, not as a logical flow. Headers, footers, page numbers, and multi-column layouts all end up jumbled when you extract text naively.

For RAG, that noise does two things: it inflates token counts without adding information, and it pollutes chunk boundaries so your retrieval returns partial or misleading segments.

A practical prep workflow

The sequence that works best for most business documents:

1. Strip metadata and hidden content first. Author names, JavaScript actions, embedded attachments, and annotations add noise and can leak into your chunks. Clean the file before extracting text.
2. Convert to Markdown. A structured Markdown extraction preserves headings, lists, and paragraph boundaries far better than plain text dumps. If the PDF has scanned pages, use OCR during this step.
3. Count tokens to pick a chunk strategy. If the full document is under your embedding model limit (say 8192 tokens for text-embedding-3-large), you may not need chunking at all. If it is over, you need to decide where to cut.
4. Chunk with overlap. Split by natural boundaries (headings or paragraphs) and add 32 to 64 tokens of overlap so retrieval does not miss information that spans a boundary.

What to preserve, what to drop

Keep: section titles, list structure, table content (even if flattened into text rows), and any reference identifiers (section numbers, clause IDs).

Drop: page numbers, repeated headers/footers, watermark text, and decorative separators. These inflate token usage without helping retrieval precision.

Tools that handle each step

If you want to run the full pipeline locally without uploading:

• Strip Hidden Content and Remove Metadata for cleanup.
• PDF to Markdown for structured extraction with OCR.
• Token Counter to measure before chunking.
• PDF Chunker for RAG to split into token-budgeted segments with overlap.

All run in the browser. The document stays on your machine through the entire pipeline.