Author(s): Sumit Vedpathak

Originally published on Towards AI.

RAG from Scratch [Part 2]: Loading — The Step Everyone Skips and Everyone Regrets

Series 2 of 5

The article argues that most RAG failures begin at the ingestion/loading stage rather than later steps like chunking or the LLM, because real-world sources (PDFs, web pages, Notion, Slack, SQL, scanned documents) are messy and often don’t load cleanly. It explains what “loading” really means—turning raw data into a clean, structured, machine-readable format—while highlighting the blurry line between parsing and loading. It then walks through loading at increasing levels of complexity: plain text, real-world file types (including PDFs with multi-column issues, OCR for scanned docs, and layout/vision-based approaches), web page extraction with content cleaning, and structured data handling by narrating CSV/SQL/JSON rows instead of losing context through naive normalization. Finally, it describes building a universal loader using frameworks that standardize outputs (text plus metadata for traceability), warns about common silent loading mistakes (bad encodings, dropped metadata, treating all PDFs the same, skipping cleaning), and closes by setting up the next series step: chunking/splitting.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI