Reducto

Reducto is a document parsing and extraction API built for AI engineering teams that need structured, LLM-ready output from complex real-world documents. Founded in 2023 by Adit Abraham and Raunak Chowdhuri, both MIT-trained ML engineers, the company graduated from Y Combinator's Winter 2024 batch and has since raised $108M in total funding from Andreessen Horowitz, Benchmark, and First Round Capital. The core problem it solves is a persistent bottleneck in AI pipelines: most documents in finance, legal, healthcare, and insurance are PDFs and spreadsheets with complex table layouts, multi-column text, embedded charts, and handwritten annotations that standard text extraction tools mangle, causing downstream model errors.

Reducto's API exposes four main endpoints. The Parse endpoint converts documents into layout-aware structured JSON blocks with bounding boxes on every element. The Extract endpoint performs schema-based structured data extraction with field-level confidence scores and provenance citations. The Split endpoint handles multi-document detection and form segmentation. The Edit endpoint, launched in 2025, is the industry's first API for editing PDFs and DOCX files dynamically without predefined templates. Supporting all of these is Reducto Studio, a no-code pipeline builder that adds confidence scoring, evaluation data generation, and pipeline performance visibility without requiring direct API integration. The platform supports 100+ languages, handles merged-cell tables, charts, handwriting, and checkboxes, and offers SOC 2 Type II, HIPAA-eligible processing with zero-retention configurations for regulated industries.

What Reducto actually does in April 2026

At its core, Reducto treats every document page as an image rather than a text file. Most PDF parsing libraries extract text from PDF metadata, which works adequately for clean, digitally created PDFs but fails on scanned documents, rotated pages, watermarked files, or any document with complex visual structure. Reducto's approach runs a three-stage pipeline: traditional computer vision analyzes the document's layout and spatial structure first; vision-language models then review and correct the initial OCR outputs in real time, a process the company calls Agentic OCR; and a final VLM pass interprets content regions contextually, linking column headers to data cells and row labels to values.

This multi-pass architecture is specifically designed for the cases where single-pass extraction silently drops data. As Adit Abraham, Reducto's CEO, explained in a September 2025 interview with AI engineer Jason Liu: "Use the best tools available..a combination of both traditional CV for things CV handles well, and also VLMs and document metadata." The practical consequence is measurable accuracy on tables that break other parsers: complex tables with merged cells spanning multiple rows and columns, nested headers where a cell's meaning depends on a hierarchy of parent headers, and tables split across page boundaries.

In November 2024, Reducto open-sourced RD-TableBench, a benchmark of 1,000 complex table images annotated by PhD-level human reviewers, covering scanned tables, multilingual content, handwriting, and multi-level header structures. This benchmark was a deliberate move to provide an objective accuracy standard in a market where vendors routinely cited incomparable internal metrics. On RD-TableBench, Reducto reports an average table similarity score of approximately 0.90, compared to AWS Textract at 0.72 and Google Cloud Document AI at 0.81.

"Despite the remarkable progress we've seen in AI over the last few years, document processing remains a critical bottleneck. Reducto exists to fix this. Unlike traditional optical character recognition methods, which.." -- @a16z, X, October 14, 2025

By October 2025, Reducto had processed over 1 billion pages, with monthly volume growing 6x between the April Series A and October Series B announcements. Named customers include Harvey, Scale AI, Mercor, Rogo, a Fortune 10 company, and a Global Top 5 Hedge Fund.

Where Reducto sits versus LlamaParse and Unstructured.io

Three competitors occupy different positions on the accuracy-versus-flexibility spectrum. Reducto is the high-accuracy, opinionated-API option. LlamaParse targets teams already using the LlamaIndex ecosystem. Unstructured.io anchors the open-source, ETL-pipeline end.

Reducto vs. LlamaParse: LlamaParse is the parsing service inside LlamaCloud, designed primarily for teams already using LlamaIndex for RAG pipeline orchestration. It supports 90+ file formats, returns multiple output formats (text, Markdown, JSON, XLSX), and has processed 500M+ documents. Where LlamaParse differs mechanically is in how it handles complex tables: it offers selectable modes (standard, LLM, agentic, agentic plus), each with different credit costs per page. Developers choose the mode manually based on expected document complexity. Reducto's Agentic OCR runs automatically and adapts per-region, which removes that configuration burden but produces higher per-page costs at the complex end. LlamaParse's schema extraction was deprecated and moved to a separate LlamaExtract service, while Reducto's Extract endpoint is integrated and produces field-level confidence scores with bounding boxes for every extracted field. For simple research-grade documents or prototyping, LlamaParse is cheaper and sufficient. For financial filings and legal contracts where a misread cell corrupts a downstream calculation, Reducto's multi-pass accuracy is worth the cost difference.

Reducto vs. Unstructured.io: Unstructured is fundamentally an open-source ETL library with an enterprise SaaS layer on top. The open-source Python package is MIT-licensed and self-hostable, using format-specific knowledge to partition documents into semantic units before chunking. Unstructured's chunking strategies (by-page, by-similarity, semantic) are more configurable than Reducto's defaults and its connector ecosystem (Databricks, Elasticsearch, S3, Google Drive, Snowflake) is broader for teams building data pipelines across multiple storage systems. The architectural tradeoff is that Unstructured's accuracy on complex tables, especially those with merged cells and nested headers in financial documents, does not match Reducto's multi-pass VLM correction. Unstructured reports an Adjusted CCT (coherent content transfer) score of 0.917 with a low hallucination-token rate, which measures different things than RD-TableBench's table similarity score. Teams prioritizing zero vendor dependency, open-source auditability, or existing ETL infrastructure integration should evaluate Unstructured. Teams prioritizing table extraction accuracy on messy real-world documents should benchmark Reducto.

"We had spent enough time exploring things that were not resonating. In comparison, it felt like we were getting punched in the face by this new idea." -- Adit Abraham, CEO, Reducto, First Round Capital interview, 2025

AWS Textract and Google Cloud Document AI also compete in this space, particularly for teams in existing AWS or GCP infrastructure. Reducto's comparison pages show 20%+ accuracy advantages over both on complex table benchmarks, though both cloud providers offer tighter infrastructure integration and simpler compliance paths for teams already within those ecosystems. For AI-native teams building purpose-built pipelines outside hyperscaler ecosystems, Reducto and Firecrawl (for web-sourced documents) tend to be the preferred ingestion pair.

What the API workflow reality looks like

Reducto's API is credit-based, billed at $0.015 per credit on the Standard pay-as-you-go tier. Standard page parsing costs 1 credit per page, so $0.015/page. Complex pages with Agentic OCR enabled cost 2 credits, putting the effective rate at $0.03/page. Additional operations stack: Agentic Extraction runs 2-4 credits per page depending on page complexity, Split costs 2 credits per page, and Edit (beta) costs 4 credits per page. A team processing 100,000 pages per month with moderate VLM usage should budget roughly $2,000-$4,000 per month before Growth tier volume discounts apply.

The Standard tier includes 15,000 free credits, which is approximately 7,500-15,000 pages depending on complexity. This covers development, testing, and early production without a credit card. The async endpoints (/parse_async, /extract_async) return immediately with a job_id and use webhook callbacks, bypassing the 200 concurrent request limit on synchronous endpoints. For high-volume batch processing, this is the recommended pattern.

Reducto Studio, the no-code pipeline builder, adds a visual layer for teams that want to configure and test pipelines without writing code. It includes confidence scoring on extracted fields, auto-generated evaluation data from real documents, and pipeline performance visibility. Up to 5 Studio seats are included on the Standard tier. At Studio's launch, users flagged crashes on iPhone Safari and Chrome due to WebGPU compatibility; the CTO acknowledged these quickly and the product has matured, but it is still newer than the core API.

Enterprise and Growth tiers add VPC deployment, on-premises options, air-gapped configurations, SSO/SAML, data residency controls, and dedicated SLAs. SOC 2 Type II and HIPAA BAA coverage is available at Enterprise. The EU note worth knowing: standard Data Processing Agreements are available on all tiers; custom redlined DPAs require Enterprise. This was a source of confusion at the Studio launch (EU developers assumed DPAs were Enterprise-only), but the default DPA covers standard GDPR requirements at Standard tier. For teams building on Pinecone or other vector databases, Reducto's structured JSON output with bounding boxes maps cleanly to the metadata schemas needed for hybrid search.

Who Reducto is built for

Reducto's customer base clusters around four verticals where document quality has direct revenue or compliance consequences. Financial services teams (hedge funds, banks, investment platforms) use it for earnings reports, 10-K filings, and financial statements where table cell misalignment can corrupt quantitative models. Legal AI platforms like Harvey use it for contracts, case filings, and regulatory submissions where field-level citation provenance is a hard requirement. Healthcare teams processing prior authorization forms and clinical documents need the checkbox detection reliability and HIPAA-eligible processing that generic VLMs cannot consistently provide. Insurance and compliance teams processing heterogeneous document batches (forms, faxes, handwritten submissions) benefit from the multilingual support and Agentic OCR's tolerance for rotated or watermarked pages.

For RAG pipelines, Reducto pairs well with AnythingLLM or LlamaIndex: bounding-box provenance on every output block enables accurate document citations, and the Extract endpoint's confidence scores let teams filter low-confidence extractions before they reach the vector index, cutting downstream hallucination rates.

AI engineering teams scaling from prototype to production commonly adopt Reducto when initial PDF parsing (PyMuPDF, pdfplumber, or naive text extraction) breaks on the 20% of their document corpus containing complex tables or scanned pages. The public RD-TableBench benchmark lets teams validate accuracy on representative samples before committing to a vendor.

What Reducto is not

Reducto is an ingestion API, not an end-to-end document workflow platform. It does not include document classification (routing invoices vs. contracts vs. forms to different pipelines), human-in-the-loop review interfaces, model fine-tuning for specific document schemas, or native integrations with downstream systems like Salesforce or SAP. Teams need to build or buy those layers separately. Competitors like Extend or Nanonets position themselves as more complete workflow solutions; Reducto's counter-argument is that a best-in-class ingestion layer that integrates with any downstream system is preferable to a vertically integrated platform with accuracy tradeoffs.

Skip Reducto when: documents are simple text-only PDFs without tables or complex layouts (LlamaParse's cost-effective mode is 5x cheaper and adequate); the team is already deep in the LlamaIndex/LlamaCloud ecosystem and wants a single vendor; open-source self-hosting with zero vendor dependency is a hard requirement (Unstructured.io); or the use case is web page scraping rather than document parsing (Firecrawl is the right tool for that job). Very early-stage teams experimenting with document AI should start with Reducto's 15K free credits before committing to a billing relationship, since the credit model can produce unexpected costs if page complexity or operation selection isn't well understood upfront.