doc-sdk: Provider-Agnostic Document SDK

📄 doc-sdk is a developer-first, ultra-lightweight TypeScript SDK that standardizes integrating visual language models (VLMs) across supported providers.

Key Features of doc-sdk

🪶 Lightweight: Tiny core, tree-shakeable, zero runtime config for the happy path.

🔌 Provider-Agnostic: Swap between ocrbase, mistral, llamaparse, and azure via a single import.

🧬 Type-Safe Extraction: First-class Zod schemas — get validated, typed JSON out of any document.

🧩 Core

• parse() — turn a document into text
• extract() — extract structured JSON from a document, with Zod-typed output
• batchParse() — batch parse
• batchExtract() — batch extract

🔌 Providers

Provider	Price ($ / 1k pages)
ocrbase	1$
Mistral OCR	2$
LlamaParse	3.75$
Azure Doc Intelligence	10–30$
AWS Textract	15–50$
Extend AI*	10–15$
Reducto*	5–10$
Unstructured*	30$

*Sales call required. Only ocrbase is available today — other providers are coming soon.

🚀 Quick Start

bun i document-sdk

Parse a document:

import { parse } from "document-sdk";

const { text } = await parse({
  file: "invoice.pdf",
});

Extract structured data:

import { extract, Output } from "document-sdk";
import { z } from "zod";

const { output } = await extract({
  file: "invoice.pdf",
  output: Output.object({
    schema: z.object({
      total: z.number(),
      vendor: z.string(),
    }),
  }),
});

Set your provider credentials in .env eg. using:

Generate ocrbase api key

OCRBASE_API_KEY=

🎛️ Picking a Provider

By default, doc-sdk auto-resolves @doc-sdk/ocrbase if installed. Override via pass a model per call:

import { parse } from "document-sdk";
import { ocrbase } from "@doc-sdk/ocrbase";

await parse({
  model: ocrbase("paddleocr"),
  file: "invoice.pdf",
});

import { parse } from "document-sdk";
import { mistral } from "@doc-sdk/mistral";

await parse({
  model: mistral("mistral-ocr-latest"),
  file: "invoice.pdf",
});