docparse-client (TypeScript)

Thin TypeScript/Node client for docparse — the pure-Rust document parser. Zero runtime dependencies: wraps the docparse binary (subprocess) or a running docparse serve (REST via global fetch). Node ≥ 18.

npm install docparse-client     # + put the `docparse` binary on PATH (or set DOCPARSE_BIN)

Parse

import { DocparseClient } from 'docparse-client';

const client = new DocparseClient();                  // finds docparse on PATH / $DOCPARSE_BIN
const doc    = await client.parse('paper.pdf');       // full IR (provenance + coordinates)
const chunks = await client.chunks('paper.pdf');      // RAG chunks: text + page + bbox + breadcrumb
const md     = await client.parse('paper.pdf', { format: 'markdown' });
const toc    = await client.outline('paper.pdf');     // structure tree (nested sections)

Against a long-running server (docparse serve --port 8642):

import { DocparseHttpClient } from 'docparse-client';
const http   = new DocparseHttpClient({ baseUrl: 'http://127.0.0.1:8642' });
const chunks = await http.chunks('paper.pdf');
const schema = await http.schema('chunk');            // the JSON Schema, for validation/codegen

Enhancements are opt-in ({ ocr, layout, tableModel, formulaModel }); digital documents never touch a model. The server must be started with the matching model config for the HTTP transport.

LangChain.js

import { DocparseLoader } from 'docparse-client/langchain';

const docs = await new DocparseLoader('paper.pdf').load();
docs[0].metadata;   // { source, page, bbox, heading_path, section_id, kind }

Every Document carries page + bbox — answers cite back to a highlightable region of the source. Needs @langchain/core (optional peer).

Vercel AI SDK

import { generateText } from 'ai';
import { docparseTools } from 'docparse-client/ai';

const result = await generateText({
  model,
  tools: await docparseTools(),     // get_chunks / outline / parse_markdown
  prompt: 'Summarize report.pdf with page citations.',
});

Needs ai + zod (optional peers; parameters-style, AI SDK v3/v4).

Types

The exported types (Chunk, Section, BBox, …) mirror the JSON contract. The authoritative schemas live in the repo (schemas/*.json, draft 2020-12) and are served at GET /openapi.json / GET /schema/{name} — run quicktype over them for exhaustive generated types.

Develop

npm install && npm run build          # → dist/
DOCPARSE_BIN=../../target/release/docparse npm test