Data ingestion from three distinct source types

Ingest at least three distinct source types (for example Markdown docs, PDF exports, and a CSV of past support tickets). Normalise everything into plain text with source metadata - title, URL, date, source type - written to ingest/corpus.jsonl, one JSON object per line.

Real knowledge bases are heterogeneous. Normalising three genuinely different formats into one shape is where you meet encoding problems, missing metadata, and silent truncation - the failures that later masquerade as "the model hallucinated".