Nutrient (formerly PSPDFKit) has launched a hosted Data Extraction API that converts PDFs, scans, images, and Office files into schema-aligned JSON with per-field citations, bounding boxes, match labels, and confidence scores. The API offers four parse modes — text ($0.001/page), structure ($0.0015/page), understand ($0.009/page), and agentic ($0.018/page) — plus auto-routing; extraction adds $0.006/page. On Nutrient's WholeDoc Bench (480 documents, 30+ languages, 7 areas, evaluated 5 November 2026), agentic mode leads with an 86.2 overall score (87.1 tables, 64.5 charts, 88.6 position) at 1.8¢/page, outperforming Claude Fable 5.1 (67.1 at 17.8¢), Azure Document Intelligence (65.9), GPT-6.1 Sol (62.9), and others. Understand mode scores 80.8 at 0.9¢/page. The API includes classify (96.2% split F1 vs. 63.7% for Typesafe AI's Jev) and split endpoints. Grounded confidence composites token probability, candidate margin, NLI grounding (Apache-2.0), and format; match labels (id_match, fuzzy_match, not_found) should be gated before confidence. Studio provides a browser workspace to configure schemas, test on documents, save versioned processors (proc_ IDs), and evaluate against up to 25 reviewed documents. Integration options include a keyless MCP server (Claude Code: `claude mcp add --transport http --scope user nutrient https://mcp.nutrient.io/mcp`), an npx-installable agent skill, a Mac app (local-only with local model), and an SDK for fully offline production. Free tier covers ~330 understand-mode extractions monthly; Starter $75/mo ($90 usage), Growth $275/mo ($345).
Read original
marktechpost