Extract tables from any PDF using dual-engine AI (Docling + pdfplumber). Low-confidence pages automatically fall back to Qwen cloud vision model.

Table Extraction Financial Verification

1 — Upload PDF

📄

Drop a PDF here, paste from clipboard, or click to browse

Max 50 MB · Financial reports, invoices, scanned tables

Shortcut: Ctrl+V / ⌘V to paste an image of a PDF from clipboard

— —

Stage: —

Progress: —

Uploading and starting extraction...

2 — Engine Settings (hover over ? for details)

AI Table Detection ?Uses IBM Docling's TableFormer ML model to find and structure tables. Disable to use only the heuristic pdfplumber engine.

Docling ML model — more accurate for complex layouts

Fast Mode ?Uses TableFormer FAST variant instead of ACCURATE. Roughly 2x speed boost on GPU, minimal accuracy loss for simple tables.

2x faster, slightly lower accuracy on complex layouts

Vision Language Model ?Replaces the entire table pipeline with SmolDocling, a 256M parameter VLM that processes the full page in one pass. Much faster on GPU; on CPU it's slower. Requires a pre-downloaded model.

SmolDocling VLM — best for scanned documents

DeepSeek High Accuracy ?Uses DeepSeek-V3 vision model as a third-pass fallback on low-confidence pages. Dramatically improves accuracy on scanned or poorly formatted PDFs. Starter & Growth only.

DeepSeek-V3 cloud vision — highest accuracy

Confidence Threshold ?Pages scoring below this threshold will be sent to Qwen cloud vision (Alibaba DashScope API) for a second opinion. Lower values save API cost but may miss poor extractions.

Pages below this score use Qwen cloud fallback

0.85

3 — Extraction Results

Total Pages

—

Tables Found

—

Qwen Fallback Calls

—

DeepSeek Fallback Calls

—

Status

—

Extracted Tables

0 cells edited — corrections will be annotated in your Excel download Reset all edits

Sign in to PDF Intelligence