Parsing Documents with LlamaIndex

Bash
pip install llama-parse

Python
from llama_parse import LlamaParse

parser = LlamaParse(
    api_key="llx-...",           # from cloud.llamaindex.ai
    result_type="markdown",      # "markdown" or "text"
    parsing_instruction="Extract all tables with full row/column structure",
    premium_mode=True,           # enables agentic/VLM parsing for complex docs
)

documents = parser.load_data("./contract.pdf")
print(documents[0].text)

For fully local/offline parsing with no LLM tokens:

Bash
npm install @llamaindex/liteparse

Progress:

• Step 1: Identify document complexity (simple text vs. scanned/handwritten/multi-modal)
• Step 2: Choose the right tool (LlamaParse cloud for complex/agentic parsing, LiteParse for fast local parsing)
• Step 3: Parse — get layout-aware markdown/text output with bounding boxes if needed
• Step 4: Extract — define a schema and run LLM-powered structured extraction
• Step 5: Split/Classify — segment long documents into logical sections, tag by natural-language rules
• Step 6: Index — chunk and embed for RAG retrieval
• Step 7: Validate — spot-check tables, charts, and handwritten sections for accuracy; re-run with premium_mode or custom parsing_instruction if quality is low

Decision guide:

• Scanned/handwritten/complex multi-column docs → LlamaParse with premium_mode=True
• Clean digital PDFs, Office docs, images, no cloud/tokens desired → LiteParse (local)
• Need structured JSON from unstructured content → LlamaParse Extract with a Pydantic schema
• Need to route document types differently → LlamaParse Classify + Split before extraction

Example 1: Input: A 40-page financial due-diligence PDF with dense multi-page tables and embedded charts. Output: Markdown output preserving table row/column relationships, chart data converted to structured JSON (labels, series, values), ready to chunk and embed for RAG.

Example 2: Input: Scanned insurance claim form with handwritten notes. Output: Clean text extraction of handwritten fields via agentic OCR auto-correction loop, structured into a defined schema (claimant name, date, claim amount, notes) with confidence flags on low-certainty fields.

Example 3: Input: Directory of 500 mixed invoices (different vendors/layouts). Output: Classified by vendor type using natural-language rules, then extracted into a unified schema (invoice_number, line_items, total) with no per-template custom code.