LlamaParse and LiteParse: IDP Software Vendor

LlamaIndex ships two complementary document parsing tools: LlamaParse, a cloud service for complex layout-aware extraction, and LiteParse, an open-source CLI released March 2026 for agent-native local parsing. They share architectural lineage but target opposite ends of the speed-accuracy tradeoff.

Platform Evolution and Market Position

LlamaParse emerged as the world's first genAI-native document parsing platform, designed specifically for LLM applications rather than traditional document workflows. Unlike legacy OCR systems that focus on text extraction, LlamaParse combines layout understanding with multimodal AI to process complex documents including charts, tables, and handwriting.

The platform has processed over 500 million documents for 300,000+ LlamaCloud users, positioning it among the most widely deployed document AI services. Built by the team behind the popular LlamaIndex framework, the service targets developers building RAG applications and AI agents that require high-quality document ingestion.

LlamaParse v2 launched in January 2026 with simplified tier-based pricing and up to 50% cost reduction, while achieving 90%+ pass-through rates versus 60-70% with legacy OCR systems through multimodal understanding and agentic workflows.

Technical Architecture and Processing Pipeline

Multimodal Processing Approach

LlamaParse processes documents through a layout-aware architecture that understands complex structures including headers, footers, and split sections. The system goes beyond text extraction to process visual context from charts, tables, and images using computer vision models. This two-stage approach of traditional extraction followed by LLM reconstruction represents a fundamental architectural difference from multi-model systems used by established players like ABBYY and Tungsten Automation.

The platform offers granular control through different parsing modes to optimize the balance between cost and accuracy. Users can input custom prompt instructions to customize output formatting, making it particularly suitable for specialized document types that defeat traditional template-based systems.

Advanced Model Integration and Performance

The platform added support for GPT-5 in preview mode with enhanced table and visual recognition capabilities, plus OpenAI's GPT 4.1 and Google's Gemini 2.5 Pro models for improved parsing accuracy. LlamaParse introduced automatic orientation and skew detection correcting documents rotated up to 270° and subtle skews between 1°-12°, plus confidence scoring ranging 0-1 for parsed pages with automatic flagging below 0.2.

Format Support and Language Coverage

Supporting 90+ document formats including PDF, PPTX, DOCX, XLSX, and HTML, LlamaParse handles both structured and unstructured content. The platform provides out-of-the-box compatibility for over 100 languages, enabling global deployment without localization requirements.

Pricing Structure and Service Tiers

LlamaParse v2 eliminated complex configuration options in favor of four standardized tiers: Fast (1 credit/page), Cost Effective (3 credits/page), Agentic (10 credits/page), and Agentic Plus (45 credits/page). The Agentic Plus tier received a 50% price reduction while maintaining equivalent accuracy.

LlamaParse operates on a freemium model with 1,000 pages per day on the free tier. The paid plan includes 7,000 pages per week plus $0.003 per additional page, making it cost-competitive with enterprise document processing platforms like Hyperscience and Instabase.

Developer Integration and Enterprise Deployment

Python SDK and API Access

The open-source Python SDK integrates directly with LlamaIndex workflows, enabling developers to process documents synchronously or asynchronously. The platform supports batch processing with configurable worker pools for high-volume applications.

from llama_parse import LlamaParse

parser = LlamaParse(
    api_key="llx-...",
    result_type="markdown",
    num_workers=4,
    verbose=True,
    language="en"
)

documents = parser.load_data("./document.pdf")

Enterprise Partnerships and Integrations

DataStax incorporated LlamaParse into RAGStack, while MongoDB partnered for Atlas Vector database integration and NVIDIA collaboration integrated with AI Enterprise platform. Davor Bonaci, CTO at DataStax, noted: "By incorporating LlamaIndex into RAGStack, we are providing enterprise developers with a comprehensive Gen AI stack that simplifies the complexities of RAG implementation."

LlamaParse offers enterprise-ready features including local cloud deployment, higher concurrency limits, and dedicated customer success support. The platform scales to process millions of pages for enterprise-grade workflows while maintaining API compatibility across deployment models.

Use Cases and Market Applications

RAG Application Development

Developers building retrieval-augmented generation systems use LlamaParse to convert messy documents into AI-ready formats. The platform's ability to preserve document structure and extract visual elements makes it particularly effective for technical documentation, scientific papers, and complex reports. Customer implementations like 11x AI's Alice SDR reducing onboarding time to days demonstrate real enterprise adoption for AI-first workflows. Developers evaluating open-source alternatives for similar RAG pipelines may also consider Unstract, which offers a no-code LLM platform with hallucination mitigation for production-grade extraction.

Financial Document Processing

Investment teams and underwriters use LlamaParse for processing invoices, insurance claims, and healthcare forms. The platform's multimodal parsing capabilities handle embedded charts and tables that traditional OCR systems often misinterpret. Eric Ciarla, co-founder at Mendable AI, emphasized: "It was easy to integrate and more powerful than any of the alternatives we tried." Teams focused specifically on structured information extraction from financial text may also evaluate LangExtract, Google's open-source Python library for LLM-powered extraction with precise source grounding. Financial analytics teams requiring document processing alongside research automation may also consider Acuity Knowledge Partners, which serves 800+ institutions with agentic AI for financial document workflows.

Platform Expansion and Future Direction

The company launched LlamaSplit Beta API for automatic document separation and LlamaExtract Table Row Mode for extracting data from repeating entities, expanding beyond traditional parsing into comprehensive document automation. This evolution from a simple PDF parser to a comprehensive document automation platform reflects the broader shift in the IDP market toward AI-native solutions.

The service's focus on RAG applications positions it in a growing niche where document processing serves as input for AI applications rather than human consumption. The platform's freemium model with 1,000 pages daily free processing and $0.003 per page pricing creates a lower barrier to entry compared to enterprise-focused competitors, potentially accelerating adoption among AI developers building RAG applications.

LiteParse: Agent-Native Local Parsing

LlamaIndex open-sourced LiteParse on March 19, 2026 as a TypeScript-native CLI and library for AI agents that need fast, reliable document parsing without cloud dependencies. Where LlamaParse maximizes accuracy through proprietary cloud models, LiteParse maximizes speed through local execution — the two products occupy opposite ends of the same parsing spectrum.

LiteParse vs LlamaParse: Which to Use

The LlamaIndex team frames the choice explicitly:

LiteParse is the core local processing engine that LlamaParse is built on, open-sourced. It does not replace LlamaParse for complex layouts or structured extraction — it extends LlamaIndex's reach into environments where cloud latency or cost is a constraint.

Why Agents Need Different Parsing

Standard document parsing tools force agents into one of two patterns: fast-but-inaccurate extraction (pypdf, pdfplumber), or slow VLM-dependent parsing that can time out during reasoning loops. Neither pattern is agent-native. Anthropic's official PDF skill encourages agents to write new Python code each time they need to process a PDF — functional but not reusable across sessions.

LiteParse is built around a two-step agent workflow: parse text for fast initial understanding, then fall back to screenshots for visual reasoning when the text output is insufficient. This pattern is explicit in the CLI design:

# Step 1: Fast text extraction for agent reasoning
lit parse report.pdf | grep "table"

# Step 2: Screenshot extraction for visual analysis
lit screenshot report.pdf -o "./report_images" --pages "1-3"

The agent skill can be installed once and reused across sessions:

npx skills add run-llama/llamaparse-agent-skills --skill liteparse

Spatial Grid Text Output

LiteParse does not attempt to detect tables and convert them to markdown. Instead, it projects text onto a spatial grid that preserves the layout relationships intact. The reasoning: LLMs are trained on ASCII tables, code indentation, and READMEs — they already understand spatial column alignment without needing explicit table markup.

Name    Age    City
John    25     NYC
Jane    30     LA

This approach eliminates an entire category of table-detection failures common in tools like PyPDF and MarkItDown, which attempt structural inference and fail when layouts don't match expected templates.

Installation and Integration

# CLI (global install)
npm i -g @llamaindex/liteparse
lit parse anything.pdf

// TypeScript/JavaScript (native library)
import { LiteParse } from '@llamaindex/liteparse';

const parser = new LiteParse({ ocrEnabled: true });
const result = await parser.parse('document.pdf');
console.log(result.text);

# Python (CLI wrapper)
# npm i -g @llamaindex/liteparse
# pip install liteparse
from liteparse import LiteParse

parser = LiteParse()
result = parser.parse("document.pdf")
print(result.text)

Format Support and OCR Pipeline

LiteParse unifies all format handling through a single PDF-centric pipeline:

OCR uses Tesseract.js built-in, which parallelizes across CPU cores automatically. For higher accuracy on difficult documents, LiteParse accepts an external OCR server:

# Built-in Tesseract.js (automatic CPU parallelization)
lit parse scanned.pdf

# External OCR server for higher accuracy
lit parse scanned.pdf --ocr-server http://localhost:8000/ocr

LlamaIndex provides example server implementations for PaddleOCR and EasyOCR. Any OCR model returning bounding boxes and text can be plugged in.

Benchmarks and Methodology

LiteParse benchmarks were conducted against PyPDF, PyMuPDF, and MarkItDown — tools targeting fast text extraction without VLM inference. The LlamaIndex team explicitly excluded VLM-based tools (LlamaParse, Docling) from benchmarks because LiteParse occupies a different performance tier.

Standard OCR benchmarks penalized LiteParse's non-markdown output and layout-relative line breaks as errors even when the text was correct. The team built a custom evaluation pipeline: LLM-generated question-answer pairs from document screenshots, manually audited, evaluated with LLM-as-a-judge. Results show LiteParse output leads to improved page-based QA accuracy versus PyPDF and PyMuPDF equivalents, with top-tier parsing latency for large documents. The dataset and evaluation code are available on HuggingFace and GitHub.

Technical Specifications

LlamaParse

LiteParse

Resources

• LlamaParse Platform
• Developer Documentation
• Python SDK
• API Sandbox
• GitHub Repository
• LiteParse Announcement
• LiteParse npm Package

Company Information

San Francisco, CA, USA Part of LlamaIndex ecosystem Founded: 2022