Knowledge Base
Choosing an engine
Several tools ship two engines: a fast deterministic path that handles clean digital documents, and a premium AI/vision path that reads the page image directly for the hard cases. Here’s how to pick.
The two-tier pattern
For tools with a vision tier, you select it with an engine parameter — auto (or tesseract for OCR) runs the deterministic pipeline; vision routes through a vision LLM that transcribes the rendered page. Vision costs more credits and is billed on the engine you request, not the path that ultimately runs.
Which tools have a vision tier
engine=autoengine=vision40 → 200engine=autoengine=vision40 → 200engine=tesseractengine=vision20 → 100 / pageengine=autoengine=vision40 → 200engine=autoengine=vision100 → 200PDF→Markdown also has a middle tier — llm_cleanup=true (100 credits) polishes the deterministic Markdown without a full vision pass.
When to use vision
- Digital, clean documents — stick with the deterministic engine. It’s faster, cheaper, and already excellent on files with a real text layer or ruled tables.
- Scanned, photographed, or degraded pages — reach for
vision. It reads the image, so it recovers numbers OCR garbles and structure the deterministic parser misses (benchmarked from near-zero to ~100% on hard scans). - Complex or borderless tables — vision wins on dense financial tables where whitespace detection struggles.
Vision tools are gated
The vision tiers pull in large models, so they require authentication — and for the AI tools (Extract Data, Summarize, Translate) a Pro/Enterprise plan. See Access tiers.Fallback behaviour
The vision engines are built to never leave you empty-handed. If a vision pass returns nothing usable — or the AI provider isn’t configured — the tool falls back to the deterministic engine rather than failing (OCR falls back page-by-page to Tesseract). The response’s engine field reports the path that actually produced the result, so you can tell when a fallback occurred, even though the credits reflect the tier you requested.