❯RESEARCH

Research

How we make answers from documents fast, grounded, and verifiable — the methods and benchmarks behind PDFLove AI.

Read the paper See benchmarks
01 · Retrieval Jun 2026

Grounded retrieval that cites every claim

  • Hybrid search
  • Reranking
  • Citations

We fuse vector and keyword retrieval with a cross-encoder reranker, then constrain generation to the retrieved passages — so every answer is traceable to a page.

Read the method
❯ Evaluated on 5,000 questions
Answer accuracy (cited)
PDFLove AI
98%
Tool A
81%
Tool B
74%
+17% vs. the next best
02 · OCR May 2026

Reading messy scans, reliably

  • OCR
  • Layout
  • 90+ languages

Our pipeline recovers text from scanned and image-only PDFs across 90+ languages, preserving tables and reading order before anything is indexed.

Read the method
❯ Benchmarked on 110 document types
Extraction F1
PDFLove AI
96.4
Tool A
88.1
Tool B
79.5
+8.3 F1 vs. the next best
Further reading
On reranking for document QA

On reranking for document QA

Jun 2026
Benchmarking OCR on real scans

Benchmarking OCR on real scans

May 2026
Citation faithfulness, measured

Citation faithfulness, measured

Apr 2026
Architecture

How a PDF becomes an answer.

From upload to cited response — OCR, chunking, hybrid retrieval, reranking, and grounded generation, each step traceable.

PDF-to-answer pipeline architecture diagram
Start free