LlamaCloud Extract (v2)

LlamaExtract turns unstructured documents (PDFs, images, text) into typed output that conforms to a schema you define. Use it when the goal is schema-conformant data for a database, dashboard, spreadsheet, or downstream model — not document text for a human or a RAG pipeline.

Is Extract the right tool?

• Yes — you know the fields you want and need them typed and validated: invoice fields, contract terms, resume attributes, table rows as records.
• No, you want raw text/markdown of the document → use llamacloud-parse.
• No, you want to search/answer questions over many documents (retrieval, RAG) → use llamacloud-index.
• Account, API key, or region setup is not covered here → core llamacloud.

Extract is built on parsing, but its contract is schema conformance and data quality, which is why schema design is where most of the work and most of the wins are.

Version posture

Extract v2 is the primary, GA path — design all new work against it. The /v1/ extract API is legacy and exists only as a migration reference; never make it the default. If you encounter v1 code, treat porting it to v2 as the goal.

Workflow

1. Confirm Extract fits (above). If the user really wants raw text or retrieval, hand off to the sibling skill.
2. Design the schema first. This dominates output quality. Apply the durable rules in references/schema-design.md: object root, shallow nesting, strong field names, descriptions that carry format/examples, optional booleans and integers, prune auto-generated fields, start minimal and iterate.
3. Choose how the schema is applied. Pick the extraction target by document shape (per-document / per-page / per-row), a tier by complexity and stakes, and decide on system prompt, page range, and citations. See references/options-and-tiers.md.
4. Run, inspect, refine. Turn citations on while tuning; fix misses by sharpening field descriptions before touching the tier.
5. Validate the result against your typed model in code and post-process anything the schema can't enforce (defaults, formats).

Schema design (the core skill)

Author the schema as a typed model in your language — Pydantic in Python, Zod in TypeScript — or hand-write JSON Schema; all compile to JSON Schema for the API. Key durable rules:

• Root is an object; wrap repeating entities in a field (arrays can't be the root).
• Keep nesting to ~3–4 levels; flatten rather than mirror the page layout.
• Names become JSON keys — choose stable, query-friendly ones.
• Descriptions guide the model — encode format and examples there.
• JSON Schema default/format are not enforced — encode them in descriptions or post-process.
• Make booleans and integers optional so "absent" isn't confused with false/0.

Full rules, the optional-field syntaxes, caps, and the split-and-merge strategy for oversized schemas are in references/schema-design.md.

Configuration (target, tier, options)

• Target by document shape: one record per document, per page, or per table row. A mismatched target is a top failure mode for repeating entities.
• Tier by complexity/stakes: the agentic default fits most documents; drop to cost-effective only after verifying accuracy; reserve the top tier for dense, high-stakes layouts.
• System prompt for global rules (disambiguation, conventions); keep field-level guidance in descriptions.
• Page range to scope large documents.
• Citations / reasoning / confidence for auditability and for diagnosing wrong extractions — the best debugging tool — at some latency cost.

Decision tables and the tuning loop are in references/options-and-tiers.md.

Minimal examples (illustrative — verify signatures live)

Python (confirm the current package name via the MCP — it has shifted between releases):

from pydantic import BaseModel
from typing import Optional

class Invoice(BaseModel):
    invoice_number: str           # field name = JSON key
    total_amount: float
    paid: Optional[bool] = None   # optional so absent != False

# Submit `Invoice` as the schema to an extraction agent, then validate the
# returned dict back through Invoice(**result) and post-process.

TypeScript (Zod):

import { z } from "zod";

const Invoice = z.object({
  invoice_number: z.string(),
  total_amount: z.number(),
  paid: z.boolean().nullable(),   // optional/absent-safe
});

These show shape, not the current API surface. Confirm the exact package name, agent/client classes, and method/parameter names via MCP (below) — the SDK package name has shifted between releases, so do not assert it from memory.

Fetching facts

Look up exact, stale-prone details live (package name, class/method signatures, parameter and enum values, the current caps, page-range syntax, pricing):

• mcp__plugin_llamacloud_docs__search_docs — concept/BM25 search.
• mcp__plugin_llamacloud_docs__read_doc — read a known doc page.
• mcp__plugin_llamacloud_docs__grep_docs — exact symbol/regex search.
• Fallback: append index.md to any https://developers.llamaindex.ai/... page URL and WebFetch it.

Anchor docs for this pillar:

• https://developers.llamaindex.ai/llamaparse/extract/
• https://developers.llamaindex.ai/llamaparse/extract/guides/schema_design/
• https://developers.llamaindex.ai/llamaparse/extract/guides/options/

For exact, current SDK signatures, prefer the MCP over any snippet in this skill.

Boundaries

• Raw text / markdown extraction → llamacloud-parse.
• Retrieval / RAG over documents → llamacloud-index.
• API keys, regions, account setup → core llamacloud.