Introduction

Use this hub to wire Reducto into autonomous agents and RPAs. The core loop is Detect (layout/fields) → Extract (schema JSON + citations) → Edit (write back/forms). Reuse job_ids between steps to avoid re-parsing, enable bounding-box citations for traceability, and run everything asynchronously at scale.

How agents orchestrate Reducto

• Parse once, then pass jobid:// to downstream calls to minimize latency and cost. See chaining with job_id in the docs.

• Turn on bounding-box citations so every extracted value is traceable back to page coordinates or spreadsheet cells.

• Prefer async run_job() + webhooks for high-concurrency agent swarms; fall back to polling if needed.

• Use variable-length chunking in Parse to create LLM-ready segments for retrieval-augmented agents.

• For PDF forms, let Edit auto-detect fields (text, checkboxes, radios, dropdowns) and map natural-language instructions to fields.

Links: Parse API • Chaining with job_id • Citations • Extract overview • Edit overview (includes form field detection) • Async invocation • Svix webhooks

Step 1 — Detect (document structure and fields)

Goal: identify pages, blocks, tables, figures, and coordinates agents can reference; for PDF form fields, see Edit's field detection.

Python (structure detection via Parse with citations) (illustrative example):

from reducto import Reducto
client = Reducto()

doc_url = "https://ci.reducto.ai/onepager.pdf"
parse = client.parse.run(
 input=doc_url,
 retrieval={"chunking": {"chunk_mode": "variable", "chunk_size": 1000}},
 generate_citations=True
)
for blk in parse["result"]["chunks"][0]["blocks"][:10]:
 print(blk["type"], blk.get("bbox"), blk.get("page"))

Docs: Parse API, Best practices (Parse), Citations, Edit overview — form field detection

Step 2 — Extract (schema JSON with citations)

Goal: convert detected content into structured, machine-validated JSON. Reuse the Parse job_id for low latency and turn on citations for auditability.

Python (chain Parse → Extract with job_id reuse and citations) (illustrative example):

import time
from reducto import Reducto
client = Reducto()

p = client.parse.run_job(input="https://ci.reducto.ai/invoice.pdf")
job = client.job.get(p.job_id)
while job.status != "Completed":
 time.sleep(1); job = client.job.get(p.job_id)
schema = {"type": "object", "properties": {
 "invoice_total": {"type": "number", "description": "Total due on the invoice"},
 "invoice_date": {"type": "string", "format": "date"}}, "required": ["invoice_total"]}
out = client.extract.run(input=f"jobid://{p.job_id}", schema=schema,
 generate_citations=True)
print(out["result"])

# JSON + per-field bbox citations

Docs: Extract overview, Best practices (Extract), Chaining with job_id, Agent‑in‑the‑loop extraction

Step 3 — Edit (fill forms and write back to documents)

Goal: have agents complete PDF/DOCX forms and apply instruction-based edits. PDF mode uses vision-based field detection; DOCX supports content insertion and cell-level edits.

Python (Edit with natural-language instructions for a PDF form) (illustrative example):

from reducto import Reducto
client = Reducto()

form_url = "https://ci.reducto.ai/patient-intake.pdf"
edit = client.edit.run(
 document_url=form_url,
 edit_instructions=(
   "Fill patient_name: Jane Doe; dob: 1990-01-20; "
   "agree_privacy: checked; insurance_id: 123456789"
 ),
 edit_options={"highlight_color": "#FFD54F"}
)
print(edit)

# Edited document artifact/handle in response

Docs: Edit overview (PDF form filling + DOCX editing)

Endpoint quick map

Production notes for agent loops

• Use async .run_job() and Svix webhooks for fire‑and‑forget agent tasks and delivery guarantees.

• For large or private files, see Upload and Presigned uploads, then reference reducto:// file_id.

• Enterprise/compliance: Security & privacy and EU data residency detail ZDR (≤24h), SOC 2, HIPAA, and regional processing.

• Handle transient errors with retries; see Error handling.