Learn RAG
Chunk size, OCR quality, PDF to Markdown, document parsing and hybrid search - what actually converts for retrieval, and what quietly does not.
What is RAG, and why does it need your documents converted?
RAG lets an AI answer from your own files instead of guessing. It works by splitting documents into passages and turning each into a vector.
RAG for PDF documents: what extracts, and what silently does not
A PDF with a text layer converts cleanly for RAG. A scanned PDF is a stack of images and yields nothing without OCR. Telling them apart matters.
RAG for Word documents: why .docx converts cleanly
A .docx is a zip of XML, so its text is already text. Word documents are among the most reliable sources for a RAG corpus.
RAG for PowerPoint: slide text converts, pictures of text do not
Slide text in a .pptx is XML and extracts cleanly for RAG. Screenshots and diagrams pasted onto slides need OCR.
RAG for Excel: why spreadsheets are the hardest easy format
Excel cell text extracts cleanly, but a grid of numbers makes poor retrieval material. What to convert, and what to leave in the database.
RAG for images and scans: why OCR is the whole job
An image carries its words as pixels, so it extracts nothing without OCR. What that means for photographs, screenshots and scanned documents.
RAG for audio and meeting recordings: transcription first
Audio needs speech recognition before it can be retrieved. Why timestamps and speaker turns matter more than transcript accuracy.
RAG for Markdown and plain text: the easiest and most overlooked
Markdown converts perfectly and its headings are ready-made chunk boundaries. Why documentation is the best first corpus.
RAG for HTML and saved web pages: everything is boilerplate
The hard part of converting web pages is discarding navigation, cookie banners and footers, which otherwise dominate the corpus.
RAG for CSV and structured data: when not to use RAG at all
Vector search over rows of numbers answers approximately what SQL answers exactly. How to decide which parts of your data belong in a RAG corpus.
Local RAG: converting documents without uploading them
Local RAG means the conversion runs on your machine. For a quick job the browser is now enough, and nothing leaves the tab.
A RAG conversion API for developers
POST a document, get JSONL back. What to expect from a document conversion API in a pipeline, and which parts are worth checking before you build on it.
Convert PDF to Markdown for LLMs: what survives, and what quietly does not
PDF to Markdown is the usual first step for LLM and RAG pipelines. What breaks is reading order, tables and scanned pages - not the text itself.
How big should a RAG chunk be?
Chunk size decides retrieval quality more than most model choices. Match it to the embedding model's real input window, not to an LLM's context length.
Does better OCR actually improve RAG retrieval?
Higher OCR character accuracy does not reliably improve RAG answers. Reading order and structure matter more than the last few percent of characters.
Document parsing for RAG: what to look for in a converter
Most document parsers for RAG differ less in what they extract than in what they do when extraction fails. Silence is the expensive failure.
Hybrid search for RAG: why keyword and vector beat either alone
Vector search finds meaning and misses exact terms. Keyword search does the opposite. Production RAG usually runs both and merges the results.
Chunking strategies for RAG: the decision that sets your ceiling
Chunk size, overlap and boundary choice affect retrieval quality more than the embedding model does. A practical guide with defaults.
What is JSONL, and why do RAG tools export it?
JSONL is one JSON object per line. It became the practical standard for moving RAG corpora because every vector database can import it.
Why convert documents locally instead of uploading them
Local conversion means contracts, medical records and internal documents never leave your machine. What that changes, and what it costs.
Embedding RAG into your application: giving AI the context it lacks
How to add retrieval to an application: where conversion fits, what to store, what it costs, and the failure modes that make a RAG feature feel unreliable.