Guides
Distill vs Extract
Which endpoint to use, when, and why
Thunderbit exposes two AI endpoints — /distill and /extract — that solve different problems. Picking the right one is the single biggest factor in cost, latency, and output quality.
When to use Distill
- You want clean, LLM-ready Markdown of an entire page
- The downstream consumer is a vector store, RAG pipeline, or LLM context
- You don't know in advance which fields you'll need
When to use Extract
- You know the exact fields you want as structured data (JSON)
- The downstream consumer is a database, dashboard, or typed code
- You want the model to do field-level reasoning (e.g. "what's the discount?")
Cost & latency tradeoffs
| Distill | Extract | |
|---|---|---|
| Credits | 1 / page | 20 / page |
| Latency | Lower (no AI extraction step) | Higher (AI step + schema validation) |
| Output | Markdown | JSON conforming to your schema |
Decision matrix
If your output is content (text, articles, knowledge base entries) → Distill. If your output is records (rows, fields, typed values) → Extract. If you're not sure, start with Distill — you can always run Extract on the markdown later.
This page is being expanded with concrete examples — check back soon.