Document Parsing & OCR Services: AWS Textract, Google Document AI, Azure Form Recognizer, Unstructured, Mathpix, Reducto, LlamaParse

⬅️ Backend & Data Overview

If you're building a B2B SaaS in 2026 that handles PDFs, scanned documents, invoices, contracts, receipts, or general unstructured content — for RAG, document understanding, expense management, contract analysis, or invoice processing — you need OCR + document parsing. The naive approach: pdf-parse npm package + regex. The structured approach: pick a service that handles tables / forms / layouts / handwriting / multi-language / complex PDFs and outputs structured data. The right pick depends on whether you need cloud-hosted-managed (AWS Textract / Google Document AI / Azure Form Recognizer) vs developer-friendly (Unstructured / Reducto / LlamaParse) vs specialized (Mathpix for math, Klippa for receipts).

TL;DR Decision Matrix

The first decision is scope: native PDF parsing (pdfplumber / PyMuPDF) vs scanned-doc OCR (Textract / Document AI / Azure) vs RAG-optimized (Unstructured / LlamaParse / Reducto). The second decision is hosting: cloud-managed vs OSS self-host vs LLM-driven (Claude / GPT-4o vision).

Decide What You Need First

Native PDF parsing — text-based PDFs (the 30% case)

Your PDFs are text-based (not scanned images). Just need text + structure extraction.

Right tools:

• pdfplumber (Python OSS) — text + tables
• PyMuPDF (Fitz) — fast PDF parsing
• pdf-parse (Node) — basic
• Unstructured — handles text + scanned

Scanned documents — OCR required (the 25% case)

PDFs are scanned images / photos. Need OCR.

Right tools:

• AWS Textract — broad; reliable
• Google Document AI — strong
• Azure Document Intelligence — alternative
• Tesseract (OSS) — DIY

Forms / structured documents (the 20% case)

Invoices, receipts, ID cards, tax forms — structured layout extraction.

Right tools:

• AWS Textract Forms / Tables
• Google Document AI Form Parser
• Azure Form Recognizer
• Klippa (receipts + invoices)
• Rossum (enterprise invoice automation)

RAG / AI pipelines (the 15% case)

Document → chunks → embeddings → vector DB. Optimized for LLM workflows.

Right tools:

• Unstructured — RAG-first; broad format support
• LlamaParse — LlamaIndex-aligned
• Reducto — modern alternative
• Anthropic Claude / OpenAI GPT-4o Vision — LLM-driven extraction

LLM-driven understanding (the 10% case)

Use Claude / GPT-4o vision to extract + reason about documents.

Right tools:

• Anthropic Claude (Sonnet 4.6 / Opus 4.7) — strong vision
• OpenAI GPT-4o — strong vision
• Vercel AI Gateway — unified across providers
• Use when extraction needs reasoning (not just text extraction)

Provider Deep-Dives

AWS Textract — broad cloud OCR

AWS managed OCR + form / table extraction.

Pricing in 2026:

• Detect Document Text: $1.50/1K pages (after 1K free)
• Forms: $50/1K pages
• Tables: $15/1K pages
• AnalyzeDocument (combined): $50/1K pages
• Custom Queries: $50/1K pages

Features: text detection, forms, tables, queries (ask questions about doc), signature detection, lending / insurance specialist models, identity documents.

Why Textract: broad capabilities; AWS-native; reliable; queries feature (ask "what's the invoice total?").

Trade-offs: pricey at scale; AWS lock-in; documents must be uploaded to S3 first.

Pick if: AWS-native; broad doc types. Don't pick if: cost-sensitive (LLM vision often cheaper now).

Google Document AI — cloud OCR + parsing

Google Cloud's document understanding.

Pricing in 2026: $1.50-30/1K pages depending on processor.

Features: OCR, form parser, invoice parser, contract parser, ID parser, receipt parser, custom parsers (train your own).

Why Document AI: Google's AI quality; specialty processors (invoice / contract / etc.); custom training for domain-specific docs.

Pick if: Google Cloud-aligned; specialty processor needed (e.g., invoice automation).

Azure Form Recognizer / Document Intelligence — cloud OCR

Microsoft's document understanding.

Pricing in 2026: free tier; ~$1.50/1K pages basic; custom $50/1K pages.

Features: prebuilt models (invoices, receipts, IDs, business cards, tax forms), custom training, layout analysis.

Why Azure: Microsoft-aligned; strong prebuilt models for common doc types.

Pick if: Azure-native; specific prebuilt model fits use case.

Unstructured — OSS / RAG-optimized

Open source + managed cloud document parser.

Pricing in 2026: OSS free (self-host); Cloud API $$ paid.

Features: 25+ format support (PDF, DOCX, PPTX, HTML, EML, etc.); chunking optimized for embeddings; metadata extraction; tables; OCR for scans (via Tesseract/Detectron2).

Why Unstructured: RAG-first; OSS option; broad format support; widely used in AI engineering.

Trade-offs: cloud API pricing not always cheaper than alternatives; OSS requires infra setup.

Pick if: building RAG pipelines; want format breadth; OSS preference. Don't pick if: simple OCR (Textract simpler).

LlamaParse — LlamaIndex-aligned

LlamaIndex's parser with LLM-enhanced extraction.

Pricing in 2026: free tier; pay-per-page paid.

Features: LLM-driven parsing (uses GPT/Claude under hood); strong on complex PDFs (academic papers, financials); markdown output.

Why LlamaParse: handles complex layouts other parsers fail on; markdown-formatted output great for LLM downstream.

Pick if: LlamaIndex-based RAG; complex PDFs (financials, academic). Don't pick if: simple OCR (overkill).

Reducto — modern document parsing

Modern alternative to Unstructured.

Pricing in 2026: pay-per-page; trial.

Features: high-quality table extraction; markdown / JSON output; LLM-friendly.

Pick if: alternative to Unstructured / LlamaParse; modern API.

Mathpix — math OCR specialist

Math / equation OCR.

Pricing: per-image.

Features: convert images of math → LaTeX / MathML; PDF math extraction.

Pick if: math-heavy documents (academic papers, math textbooks). Don't pick if: general docs.

Klippa — receipt / invoice specialist

Receipt + invoice OCR.

Pricing: per-doc.

Features: receipt OCR with line-item extraction; invoice extraction; expense categorization.

Pick if: expense management / receipt automation product. Don't pick if: general docs.

Tesseract — OSS OCR

Open-source OCR engine.

Pricing: free; you host.

Features: 100+ languages; handwriting OK-ish; baseline OCR.

Pick if: privacy-sensitive; cost-priority; can self-host.

Trade-offs: lower accuracy than commercial cloud OCR; no built-in form / table extraction.

PaddleOCR — OSS OCR (Baidu)

OSS alternative to Tesseract.

Features: strong multilingual (Chinese especially); table extraction; layout analysis.

Pick if: multilingual / Chinese-heavy.

LLM Vision (Claude / GPT-4o) — LLM-driven OCR

Use vision LLMs for document understanding.

Pricing in 2026:

• Claude Sonnet 4.6: ~$3/1M input + $15/1M output tokens
• Claude Opus 4.7: ~$15/1M input + $75/1M output tokens
• GPT-4o: ~$2.50/1M input + $10/1M output tokens

Features: vision + reasoning combined; ask questions; structured output; complex layouts.

Why LLM vision: handles edge cases other OCR misses; can extract + reason in one call; markdown output.

Trade-offs: cost varies (often cheaper than Textract for low volume; expensive at high volume); rate limits; hallucination risk.

Pick if: complex extraction with reasoning; low-medium volume. Don't pick if: high volume + strict accuracy (calibrate carefully).

ABBYY / Rossum / Nanonets — enterprise

Enterprise-focused OCR.

Pricing: custom; $$.

Pick if: enterprise procurement; specific use case (invoice automation).

What Document Parsing Won't Do

Buying a parsing service doesn't:

1. Solve all OCR errors. ~95-99% accuracy on common types; 80-90% on edge cases (handwriting, low-quality scans). Plan for review.
2. Replace human-in-the-loop for legal docs. Contracts / legal docs benefit from review even with high-accuracy OCR.
3. Handle every layout perfectly. Tables nested in tables, multi-column with sidebars, fillable forms with markup — edge cases break parsers.
4. Embed semantic understanding. OCR extracts text; understanding requires LLM downstream.
5. Process images at acceptable speed for real-time UX. Many APIs are seconds-to-minutes per doc; plan async.

The honest framing: document parsing is statistical. 95% works fine for most. The 5% will surprise you.

LLM Vision vs Traditional OCR — 2026 Decision

Decide LLM vision vs traditional OCR.

Use LLM vision (Claude / GPT-4o) when:
- Complex layouts (multi-column, sidebars, mixed content)
- Need reasoning + extraction in one call ("what's the total?")
- Handwriting / poor quality scans
- Low-medium volume (<10K docs / month)
- Cost is OK ($0.01-0.10 per doc with vision)

Use traditional OCR (Textract / Document AI) when:
- Simple layouts; high accuracy required
- High volume (>100K docs / month) — cost difference matters
- Specific specialty model fits (invoice, receipt, ID)
- Compliance requires deterministic processing
- Existing AWS / Google / Azure stack

Use OSS (Tesseract / PaddleOCR) when:
- Privacy critical (no cloud)
- Cost dominant; can absorb engineering
- Self-host expertise on team

Hybrid pattern (recommended for many):
- Tier 1: traditional OCR for fast / high-confidence extraction
- Tier 2: LLM vision for edge cases + reasoning
- Tier 3: human review for low-confidence

Decision factors:
- Volume × per-doc cost
- Accuracy needs
- Latency requirements
- Privacy / compliance

Output:
1. Recommendation for [USE CASE]
2. Per-doc cost calculation
3. Latency / throughput estimate
4. Failure-mode handling
5. Migration path if scale changes

The 2026 trend: LLM vision is becoming primary for low-volume + complex docs. Traditional OCR remains better for high-volume + simple. Many teams blend.

Pragmatic Stack Patterns

Pattern 1: Indie SaaS adding doc upload ($0-100/mo)

• Unstructured OSS (self-host) OR Tesseract
• Or LLM vision (Claude / GPT-4o) for low volume
• Total: <$100/mo

Pattern 2: SMB B2B SaaS with PDF features ($100-2K/mo)

• AWS Textract OR Unstructured Cloud OR LlamaParse
• 1K-100K docs / mo
• Total: $100-2K/mo

Pattern 3: Mid-market with structured docs ($1K-10K/mo)

• Google Document AI with custom processors
• Or Azure Form Recognizer
• Custom training for domain
• Total: $1-10K/mo

Pattern 4: RAG / AI pipelines ($500-5K/mo)

• Unstructured OR LlamaParse OR Reducto
• Optimized for chunking + embeddings
• Vector DB integration

Pattern 5: LLM-driven extraction ($cheap-expensive)

• Anthropic Claude OR OpenAI GPT-4o
• Via Vercel AI Gateway for fallback
• Pay-per-token; varies wildly with volume

Pattern 6: Receipt / invoice product ($per-doc)

• Klippa OR Rossum OR Google Document AI Invoice Parser
• Specialty model

Pattern 7: Math / academic ($per-doc)

• Mathpix
• Or LLM vision for general academic

Pattern 8: Privacy / OSS ($hosting)

• Tesseract OR PaddleOCR OR Unstructured OSS
• Self-host
• Total: VPS cost only

Decision Framework: Three Questions

1. What kind of documents?

   • Native PDF (text-based) → pdfplumber / PyMuPDF / Unstructured
   • Scanned / images → Textract / Document AI / Azure / Tesseract
   • Forms (invoice / receipt / ID) → Document AI / Form Recognizer / Klippa
   • Math / equations → Mathpix
   • General + reasoning → LLM vision (Claude / GPT-4o)

2. What's your volume?

   • <1K docs/mo → free tiers / LLM vision
   • 1K-100K → cloud OCR (Textract / Document AI)
   • 100K+ → negotiate enterprise; consider OSS for cost

3. What's downstream?

   • RAG → Unstructured / LlamaParse / Reducto
   • Form data extraction → Document AI / Form Recognizer
   • Search index → standard OCR
   • LLM understanding → LLM vision

Verdict

For 30% of B2B SaaS in 2026 needing doc parsing: Unstructured (OSS or cloud) — RAG-first; broad format.

For 25%: AWS Textract — AWS-native; broad capability.

For 15%: Google Document AI — Google Cloud + specialty processors.

For 10%: LLM vision (Claude / GPT-4o) — low-medium volume + reasoning.

For 10%: LlamaParse / Reducto — RAG modern.

For 5%: Klippa / Rossum — receipts / invoices.

For 5%: Tesseract / PaddleOCR / pdfplumber — OSS / privacy.

The mistake to avoid: expecting 100% accuracy. Plan for 95% and human review for low-confidence. Surface confidence scores to users.

The second mistake: using LLM vision at high volume without cost analysis. $0.05/doc × 1M docs = $50K/month. Math first.

The third mistake: not testing on your actual documents. Sample 50 real docs; run through 3 services; measure accuracy. Marketing demos are misleading.

See Also

• Vector Databases — RAG storage downstream
• PDF & Document Generation Tools — flip side (generating PDFs)
• Unstructured — Unstructured deep-dive
• Search Providers — adjacent search infrastructure
• Notification Providers — adjacent infra
• API Integration — generic integration patterns
• VibeReference: Anthropic Claude — Claude vision
• VibeReference: OpenAI GPT — GPT-4o vision
• VibeReference: AI Gateways — unified routing
• VibeReference: Vercel AI Gateway — Vercel-managed
• VibeWeek: RAG Implementation — full RAG pipeline
• VibeWeek: AI Features Implementation — AI feature patterns
• VibeWeek: File Uploads — upload pipeline upstream
• VibeWeek: Image Upload Processing Pipeline — adjacent
• VibeWeek: Markdown Rendering & Sanitization — render parsed output