[01/08] Ingest • Reading the source files
Step zero — the system reads your PDF
Before anything smart happens, the pipeline extracts every word from the policy. You'll meet 5 sample pages here and in every demo (1, 2, 8, 12, 31). Garbage in → garbage out, so this step decides answer quality.
How it works
- PDF pages go in — scanned or digital.
- Text and tables are extracted (OCR model only if scanned).
- Every slice's Meaning Fingerprint gets stored, ready for Search.
What gets extracted — tap a page
Page 1 — Definitions (digital text)
Page 8 — Provider table
Page 31 — Scanned form (needs OCR)
Tap a page above to see how it is read.
What it does
- Reads text, tables and images (OCR)
- Keeps data inside your system — private
- Acts as the single source of truth
What it does NOT do
- No slicing yet (that’s Chunking)
- No meaning yet (that’s Embeddings)
- Does not “look” at the PDF like you do — it extracts
Next: Chunking
Extracted text is still 50 pages. Next the pipeline cuts it into slices →