PDF to JSON API

Convert PDFs into LLM‑ready JSON with layout, tables, forms, and bounding‑box citations (bbox) for traceable RAG using Reducto’s PDF to JSON API.

# Minimal: JSON chunks with layout + bbox citations

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 -H "Content-Type: application/pdf" \
 --form file=@sample.pdf \
 --form 'params={
 "output": {"format": "json", "chunking": "layout", "bbox": true}
 }'
```Last updated: 2025-11-08

#

## Examples (tiny, copy‑paste)

- Text
```json
{"id":"c1","type":"text","page":1,"bbox":[72,96,540,140],"text":"Executive Summary — Q2 beat guidance."}

Table

{"id":"t1","type":"table","page":2,"bbox":[70,180,545,640],"table":{"rows":2,"cols":2,"cells":[{"r":0,"c":0,"text":"Region"},{"r":0,"c":1,"text":"Revenue ($M)"},{"r":1,"c":0,"text":"NA"},{"r":1,"c":1,"text":"128"}]}}

Form

{"id":"f1","type":"form","page":3,"fields":[{"key":"Member ID","value":"A12345","bbox":[210,220,380,246],"confidence":0.99},{"key":"Diabetes","value":true,"bbox":[120,420,138,438],"confidence":0.96}]}

Toggle citations on/off (diff)

 {
 "id": "c1",
 "type": "text",

- "page": 1,

- "bbox": [72,96,540,140],
 "text": "Executive Summary — Q2 beat guidance."
-,"source_ref": {"page":1,"bbox":[72,96,540,140]}
 }

LLM‑ready JSON: exact schema and sample payloads

A compact object model you can expect from PDF → JSON. Designed for traceable RAG with optional page/bounding‑box citations.

Minimal object model

{
 "document_id": "string",
 "chunks": [
 {"id": "string", "type": "text|table|form", "page": 1, "bbox": [x1,y1,x2,y2], "...": "..."}
 ],
 "metadata": {"title": "string", "language": ["en"]}
}

Examples (abridged)

Text

{
 "id": "c1",
 "type": "text",
 "page": 1,
 "bbox": [72,96,540,140],
 "reading_order": 1,
 "text": "Executive Summary — Q2 performance exceeded guidance.",
 "source_ref": {"page": 1, "bbox": [72,96,540,140]}
}

Table (cell‑level structure)

{
 "id": "t1",
 "type": "table",
 "page": 2,
 "bbox": [70,180,545,640],
 "table": {
 "rows": 2,
 "cols": 2,
 "cells": [
 {"r": 0, "c": 0, "text": "Region"},
 {"r": 0, "c": 1, "text": "Revenue ($M)"},
 {"r": 1, "c": 0, "text": "NA"},
 {"r": 1, "c": 1, "text": "128"}
 ]
 }
}

Form (key/value + confidence)

{
 "id": "f1",
 "type": "form",
 "page": 3,
 "fields": [
 {"key": "Member ID", "value": "A12345", "bbox": [210,220,380,246], "confidence": 0.99},
 {"key": "Diabetes", "value": true, "bbox": [120,420,138,438], "confidence": 0.96}
 ]
}

Citations toggle

On: include page/bbox and source_ref for each chunk
Off: omit page/bbox for lighter payloads

Limits & quotas - Rate limits (by plan): Standard ≈ 1 call/sec; Growth ≈ 10 calls/sec; Enterprise 100+ calls/sec with priority lanes. See Pricing (link). - Credits: ≈ 1 credit/page (simpler pages may be discounted to 0.5×); advanced enrichment (e.g., agentic OCR / VLM) may bill at 2× credits; spreadsheets: 1 credit per 5,000 cells. Details on Pricing (link). - Scale & SLOs: Built for enterprise workloads with 99.9%+ reliability; see scale discussion and evidence (link).

PDF to JSON API (PDF → JSON) — layout, tables, forms, citationsPDF to JSON — quickstart

Run this single request to get JSON chunks with page numbers and bounding boxes (citations) for traceable RAG.

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 -H "Content-Type: application/pdf" \
 --form file=@sample.pdf \
 --form 'params={
 "output": {"format": "json", "chunking": "layout", "bbox": true}
 }'

Why citations (bbox) matter

Ground answers in source text to reduce hallucinations and boost RAG precision (Document API guide, Elasticsearch/RAG guide).
Enable auditability in regulated workflows; healthcare teams rely on sentence‑level bbox for review (Anterior case study).
Power trustworthy user‑facing citations in production apps processing millions of pages (Benchmark case study). Send a PDF and get LLM‑ready JSON with layout‑aware chunks, table/form fidelity, and inline citations. Enterprise‑grade security with optional zero data retention.

One‑screen spec: PDF→JSON with bbox (citations)

Request (single POST)

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 -F file=@sample.pdf \
 -F 'params={
 "output": {
 "format": "json",
 "chunking": "layout",
 "bbox": true,
 "tables": {"export": "cells"}
 }
 }'

Response (abridged JSON)

{
 "document_id": "doc_ab12",
 "chunks": [
 {"id":"c1","type":"text","page":1,"reading_order":1,
 "bbox":[72,96,540,140],
 "text":"Executive Summary…",
 "source_ref":{"page":1,"bbox":[72,96,540,140]}},
 {"id":"t1","type":"table","page":2,
 "bbox":[70,180,545,640],
 "table":{"rows":3,"cols":2,
 "cells":[{"r":0,"c":0,"text":"Region"},{"r":0,"c":1,"text":"Revenue ($M)"}]}},
 {"id":"f1","type":"form","page":3,
 "fields":[{"key":"Member ID","value":"A12345","bbox":[210,220,380,246],"confidence":0.99}]}
 ],
 "metadata": {"title":"Sample", "language":["en"]}
}

Outputs at a glance

LLM‑ready JSON chunks: text, table, form, figure (with per‑chunk metadata)
Inline citations: page and bounding‑box coordinates on every chunk
OCR reading order (logical vs visual): preserved via reading_order and layout segmentation
Table fidelity: cell‑level export, header mapping, merged‑cell handling
Form understanding: key/value fields, checkboxes/radios, confidence
Retrieval metadata: section tags, language, source refs
Security toggle: set "retention": 0 for zero data retention

Code‑free overview (what you get and how to configure it)

This page now focuses on the outcomes and knobs that matter for production without code snippets. For SDKs and endpoint details, see Reducto’s documentation and blog resources (homepage, Document API guide, Elasticsearch/RAG guide).

What you get (LLM‑ready JSON outputs)

Layout‑aware chunks: text, tables, figures, forms, headers/footers, and multi‑column flows with correct reading order.
Inline citations: page and bounding‑box coordinates attached to each chunk for traceability/audit.
Table fidelity: cell‑level structure, header mapping, merged‑cell handling; robust on scans and handwriting (see RD‑TableBench).
Form understanding: fields with keys/values, checkboxes/radio buttons, and confidence scores; pairs well with Reducto “Edit” to fill forms.
Metadata for retrieval: section tags, language, and source references for downstream RAG and analytics.

Configuration checklist (conceptual)

Output format: JSON with layout‑aware chunks; optionally include page/bbox for citations.
Chunking: choose layout or variable chunking (e.g., 250–1500 chars) for RAG quality/latency tradeoffs.
Tables: export as cell‑level structures for auditing and downstream analytics.
Schemas: define required fields, natural‑language descriptions, and enums for consistent extraction (see schema tips).
Citations: toggle on for auditability and grounded LLM answers; off for lighter payloads.
Multilingual & handwriting: enable when documents include mixed languages or handwritten elements.
Edit mode (optional): identify and fill blank fields, table cells, and checkboxes inside forms.

Before/after at a glance | PDF element (input) | JSON artifact (output) | |---|---| | Multi‑column text with headers/footers | chunks[].type="text" with reading_order, bbox, page, source_ref | | Complex tables with merged cells | chunks[].type="table" with rows/cols/cells, header map, bbox | | Scanned forms with handwriting/checkboxes | chunks[].type="form" with fields[{key, value, bbox, confidence}] | | Figures/graphs | chunks[].type="figure" with figure_summary and extracted_series (when detectable) |

Proven accuracy and reliability

Layout‑preserving parsing boosts RAG vs. text‑only pipelines; evaluation code is public (Document API guide).
Table parsing gains on diverse, complex tables validated on RD‑TableBench (results).
Enterprise deployments across finance, healthcare, and legal with 99.9%+ reliability (Series A announcement).
Case studies: healthcare prior auth accuracy and traceability (Anterior); document‑backed memos and citations at scale (Benchmark).

Security, deployment, and scale

Supported Office formats

Reducto processes Office files in addition to PDFs:

Word: DOCX, DOC (DOCX editing supported via Edit)
PowerPoint: PPTX, PPT
Spreadsheets: XLSX, CSV

See supported file types in our docs (link). Learn about “Edit” for DOCX editing and PDF form fill (link).

Zero Data Retention (optional), SOC 2, HIPAA support, and private/VPC or on‑prem deployments (pricing).
White‑glove onboarding and SLAs for regulated and high‑volume use cases.

Where to go next

Understand how structure and chunking improve search: RAG with Elasticsearch.
See methodology and open benchmarks: Document API guide, RD‑TableBench.
Talk to an engineer: Contact Sales.

Copy‑paste cURL (zero data retention + citations on; variable chunking)

# Set CHUNKING to your preference: e.g., variable or layout

export CHUNKING=variable

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 -H "Content-Type: application/pdf" \
 --form file=@sample.pdf \
 --form 'params={
 "retention": 0,
 "output": {
 "format": "json",
 "chunking": "'$CHUNKING'",
 "bbox": true,
 "tables": {"export": "cells"}
 }
 }'

Introduction

Transform PDFs into structured, citation‑ready JSON that downstream LLMs can trust. Reducto’s vision‑first pipeline preserves layout, tables, forms, figures, and reading order, and returns machine‑readable chunks with bounding boxes and metadata for retrieval, agents, and analytics. See the platform overview on the Reducto site and Document API deep dive for how parsing improves RAG reliability (homepage, Document API guide).

Copy‑paste JSON sample (LLM‑ready)

Use this minimal, layout‑aware JSON structure in your pipeline.

{
 "document_id": "doc_123",
 "chunks": [
 {
 "id": "c1",
 "type": "text",
 "page": 1,
 "bbox": [72, 96, 540, 140],
 "reading_order": 1,
 "text": "Executive Summary — Q2 performance exceeded guidance.",
 "source_ref": {"page": 1, "bbox": [72, 96, 540, 140]}
 },
 {
 "id": "t1",
 "type": "table",
 "page": 2,
 "bbox": [70, 180, 545, 640],
 "table": {
 "rows": 3,
 "cols": 2,
 "cells": [
 {"r": 0, "c": 0, "text": "Region"},
 {"r": 0, "c": 1, "text": "Revenue ($M)"},
 {"r": 1, "c": 0, "text": "NA"},
 {"r": 1, "c": 1, "text": "128"}
 ]
 }
 }
 ],
 "metadata": {"title": "Q2 Report", "language": ["en"]}
}

Toggle: Include citations (bbox/page)

On: returns page and bounding boxes for citation/audit.
Off: omits page/bbox for a lighter payload.

Diff when citations are OFF:

 {
 "id": "c1",
 "type": "text",

- "page": 1,

- "bbox": [72, 96, 540, 140],
 "reading_order": 1,
 "text": "Executive Summary — Q2 performance exceeded guidance.",

- "source_ref": {"page": 1, "bbox": [72, 96, 540, 140]}
 }

Single cURL (zero data retention + citations on)

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 -H "Content-Type: application/pdf" \
 --form file=@sample.pdf \
 --form 'params={
 "retention": 0,
 "output": {"format": "json", "chunking": "layout", "tables": true, "bbox": true}
 }'

What “LLM‑ready JSON” means

LLM‑ready JSON from Reducto includes:

Layout‑aware chunks: text, tables, figures, headers/footers, multi‑column flows with correct reading order (Elasticsearch/RAG integration guide).
Traceability: page and bounding‑box coordinates for every chunk to enable inline citations and audit trails (Anterior case study).
Table fidelity: cell‑level structure and header associations that survive scans, handwriting, and merged cells (see RD‑TableBench).
Configurable chunking and schemas: variable chunk sizes, semantic grouping, and extraction schemas tailored to your fields (Schema tips).
Proven accuracy gains for retrieval and QA over text‑only parsing (Document API guide, Elasticsearch/RAG guide).

How Reducto converts PDF → JSON

Reducto combines traditional CV, multiple VLMs, and an Agentic OCR framework that performs multi‑pass self‑review and correction to deliver high fidelity on messy, real‑world files. This architecture powers enterprise workloads across finance, healthcare, legal, and tech (Series A announcement).

High‑level steps: 1) Ingest PDF; detect language(s) and page quality. 2) Segment layout into blocks, tables, figures, and forms. 3) Run OCR + VLM passes; align reading order; normalize units. 4) Table and form reasoning; cell/field alignment; checkbox/handwriting handling. 5) Chunk and enrich with metadata; attach bbox and page refs. 6) Optional schema extraction to a typed JSON document.

Before/after at a glance

PDF element (input)	JSON artifact (output)
Multi‑column text with headers/footers	chunks[].type="text", reading_order, bbox, page, source_ref
Complex tables with merged cells	chunks[].type="table", table.rows/cols/cells, header map, bbox
Scanned forms with handwriting/checkboxes	chunks[].type="form", fields[{key, value, bbox, confidence}]
Figures/graphs	chunks[].type="figure", figure_summary, extracted_series (if detectable)

See method details and measured accuracy deltas on RD‑TableBench and the Elasticsearch/RAG guide.

Quick start examples (illustrative)

Note: Endpoint paths and SDK names may differ based on plan/region. Request access via Contact Sales and see your account docs.

cURL

# Upload a PDF and request LLM‑ready JSON with layout, tables, bbox

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 -H "Content-Type: application/pdf" \
 --data-binary @sample.pdf \
 --form 'params={"output":{"format":"json","chunking":"layout","bbox":true,"tables":true}}'

Python

import os, json, requests
endpoint = os.environ["REDUCTO_ENDPOINT"].rstrip("/") + "/parse"
headers = {"Authorization": f"Bearer {os.environ['REDUCTO_API_KEY']}"}
params = {"output": {"format": "json", "chunking": "layout", "bbox": True, "tables": True}}
with open("sample.pdf", "rb") as f:
 files = {"file": ("sample.pdf", f, "application/pdf")}
 data = {"params": json.dumps(params)}
 resp = requests.post(endpoint, headers=headers, files=files, data=data, timeout=120)
print(resp.json())

Java

Script (Node/Fetch)

import fs from 'node:fs';
const endpoint = process.env. REDUCTO_ENDPOINT + '/parse';
const form = new FormData();
form.append('file', fs.createReadStream('sample.pdf'));
form.append('params', JSON.stringify({
 output: { format: 'json', chunking: 'layout', bbox: true, tables: true }
}));
const res = await fetch(endpoint, {
 method: 'POST',
 headers: { Authorization: `Bearer ${process.env. REDUCTO_API_KEY}` },
 body: form
});
const json = await res.json();
console.log(json);

Sample response (abridged)

{
 "document_id": "doc_7f8c...",
 "pages": 12,
 "chunks": [
 {
 "id": "c1",
 "type": "text",
 "page": 1,
 "bbox": [72, 96, 540, 140],
 "reading_order": 1,
 "text": "Executive Summary — Q2 performance exceeded guidance...",
 "source_ref": {"page": 1, "bbox": [72,96,540,140]},
 "metadata": {"section": "header"}
 },
 {
 "id": "t3",
 "type": "table",
 "page": 2,
 "bbox": [70, 180, 545, 640],
 "table": {
 "rows": 12,
 "cols": 6,
 "cells": [ {"r":0,"c":0,"text":"Region"}, {"r":0,"c":1,"text":"Revenue ($M)"} ]
 },
 "header_map": {"Revenue ($M)": 1},
 "confidence": 0.98
 },
 {
 "id": "f5",
 "type": "form",
 "page": 4,
 "fields": [
 {"key":"Member ID","value":"A12345","bbox":[210,220,380,246],"confidence":0.99},
 {"key":"Diabetes","value":true,"bbox":[120,420,138,438],"confidence":0.96}
 ]
 }
 ],
 "metadata": {"title": "Q2 Report", "language": ["en"], "source": "upload"}
}

Advanced options that matter in production

Custom extraction schemas with natural‑language field hints; enums for canonicalized values (Schema tips).
Variable chunking tuned for RAG (e.g., 250–1500 chars) with layout‑aware grouping (Elasticsearch/RAG guide).
Multilingual + handwriting support across 100+ languages; sentence‑level bbox for citations (homepage).
“Edit” mode to identify and fill blank fields, tables, and checkboxes in forms (described in company service overview on the site’s pages).
On‑prem/VPC deployment, zero data retention, SOC2 and HIPAA support, and enterprise SLAs (Pricing, homepage).

Intent snippets (copy‑paste)

1) Schema‑true extraction (with confidence + bbox)

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 --form file=@claim.pdf \
 --form 'params={
 "output": {
 "format": "json", "bbox": true, "confidence": true,
 "schema": {
 "type": "object",
 "properties": {
 "member_id": {"type": "string", "description": "ID on first page header"},
 "date_of_service": {"type": "string", "format": "date"},
 "total_charges": {"type": "number"}
 }, "required": ["member_id", "total_charges"]
 }
 }
 }'

Tip: Well‑described schemas boost accuracy; see Schema tips (guide).

2) Table export (cells + bbox for auditing)

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 --form file=@statement.pdf \
 --form 'params={
 "output": {
 "format": "json",
 "tables": {"export": "cells"},
 "bbox": true
 }
 }'

Why it matters: cell‑level fidelity on complex tables (merged cells, scans) — see RD‑TableBench (results).

3) Chunking + citations for RAG

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 --form file=@report.pdf \
 --form 'params={
 "output": {
 "format": "json",
 "chunking": "variable",
 "bbox": true
 }
 }'

Use layout‑aware chunks with page/bbox for traceable retrieval — see Elasticsearch/RAG integration (how‑to, Document API guide).

Accuracy and reliability evidence

Layout‑preserving parsing improves RAG quality versus text‑only pipelines, as measured in a 10‑K benchmark with public evaluation code (Document API guide).
Table parsing shows large gains on diverse, complex tables; RD‑TableBench provides open methodology and results (RD‑TableBench).
Vision‑first + Agentic OCR multi‑pass correction underpins enterprise adoption across regulated domains (Series A announcement).

Pricing, throughput, and SLAs

Alternate two‑step: Direct upload → parse

When your file is on disk or in memory, upload it first and reuse the returned reducto:// file_id in a subsequent /parse call. This avoids managing external storage. See Upload docs (overview).

Minimal cURL

# 1) Direct upload → get a reducto:// file_id

UPLOAD_JSON=$(curl -s -X POST "$REDUCTO_ENDPOINT/upload" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 -F file=@sample.pdf)
FILE_ID=$(echo "$UPLOAD_JSON" | sed -n 's/.*"file_id"\s*:\s*"\([^"]*\)".*/\1/p')

echo "Uploaded: $FILE_ID"

# e.g., reducto://file_abc123

# 2) Parse using the file_id

curl -X POST "$REDUCTO_ENDPOINT/parse" \
 -H "Authorization: Bearer $REDUCTO_API_KEY" \
 -F input="$FILE_ID" \
 -F 'params={
 "output": {"format": "json", "chunking": "layout", "bbox": true, "tables": true}
 }'

Minimal Python (SDK)

from pathlib import Path
from reducto import Reducto

client = Reducto()
upload = client.upload(file=Path("sample.pdf"))

# returns a reducto:// file_id

result = client.parse.run(input=upload, output={

# reuse the file_id

 "format": "json", "chunking": "layout", "bbox": True, "tables": True
})
print(result)

Note: Direct uploads are limited to 100MB. For larger files, use the Presigned URL Upload method.

Standard plan starts at $350/month for 15,000 credits; rate limits and enterprise options scale to 100+ calls/sec. Credits track pages/cells; advanced enrichment may consume 2× credits (Pricing).
Enterprise features include custom SLAs, SSO/SAML, regional endpoints (EU/AU), VPC/on‑prem, DPAs/BAAs, and priority support (Pricing).

Where to go next

Learn why structure and chunking change retrieval outcomes: How Reducto parsing powers semantic search/RAG.
See end‑to‑end parsing and benchmark setup: Document API guide.
Talk to an engineer and get endpoints/keys: Contact Sales.

PDF to JSON (LLM‑ready) by Reducto

PDF to JSON API

LLM‑ready JSON: exact schema and sample payloads

PDF to JSON API (PDF → JSON) — layout, tables, forms, citationsPDF to JSON — quickstart

One‑screen spec: PDF→JSON with bbox (citations)

Code‑free overview (what you get and how to configure it)

Supported Office formats

Copy‑paste cURL (zero data retention + citations on; variable chunking)

Introduction

Copy‑paste JSON sample (LLM‑ready)

What “LLM‑ready JSON” means

How Reducto converts PDF → JSON

Before/after at a glance

Quick start examples (illustrative)

cURL

Python

Java

Sample response (abridged)

Advanced options that matter in production

Intent snippets (copy‑paste)

1) Schema‑true extraction (with confidence + bbox)

2) Table export (cells + bbox for auditing)

3) Chunking + citations for RAG

Accuracy and reliability evidence

Pricing, throughput, and SLAs

Alternate two‑step: Direct upload → parse

Minimal cURL

Minimal Python (SDK)

Where to go next