Week 6 · Enterprise AI

Tables break text splitters

0:00 / 3:00

15 seconds

The dose landed in another chunk.

A dosing table was split mid-row. The model quoted the drug with the wrong milligrams.

PDF tableBroken rowsMarkdownIntact tableChunksKeep rows
Lesson notesTables break text splitters: the short version+

Week 6 · RAG and evals · about 3 minutes

What you will be able to do

Convert Helix policy tables to Markdown before chunking so rows stay together.

The idea

Character splitters ignore table structure. A row can be cut between drug and dose.

Convert tables to Markdown (or HTML) first, then chunk on headings or whole tables so a row stays one unit.

| Drug | Dose |
|------|------|
| Aspirin | 81 mg |
| Warfarin | per INR |

If the table matters clinically, never split through a row.

Key takeaways

  • Tables → Markdown (or keep whole).
  • Chunk on headings / table units.
  • Never bisect a clinical row.

Part of the Forward Deployed Engineer roadmap. New to the role? Start with what a Forward Deployed Engineer does.