Introduction

Reducto is the complete agentic document platform for leading AI teams needing performance at enterprise scale. The API is engineered for sustained, elastic throughput on real‑world documents with a 99.9% uptime target, automatic scaling, and custom SLAs for Enterprise plans. See the platform overview and reliability notes on the Reducto site and blog: Reducto platform, RAG at enterprise scale, and Pricing (plan SLAs and rate limits).

Plan rate limits (QPS)

Rate limits are plan‑scoped and can be raised for Enterprise. Reference: Pricing.

Notes: “pages/min” examples assume one page per request (PPR=1). Actual throughput scales with average pages per request and document mix. See credit/complexity details in Credit usage.

QPS → pages/min calculator

Use this deterministic conversion to estimate sustained ingest capacity.

• pages_per_minute = QPS × 60 × average_pages_per_request (PPR)

• Example 1 (Growth): 10 QPS × 60 × 3 PPR ≈ 1,800 pages/min.

• Example 2 (Enterprise): 100 QPS × 60 × 2 PPR ≈ 12,000 pages/min.

• Example 3 (Standard): 1 QPS × 60 × 5 PPR ≈ 300 pages/min.

Important context:

• Complexity affects credits billed (e.g., agentic/VLM pages cost more credits) and may influence processing latency by document mix; see Credit usage.

• Large responses may be returned via presigned URLs; see the /parse reference.

Batch queue: 20% off parsing for non-urgent volume

For workloads where minutes-to-hours latency is acceptable — backfills, archive migrations, overnight runs — Reducto offers a discounted batch lane: submit /parse_async jobs with queue_priority: "batch" and receive a 20% credit discount on parsing, with completion guaranteed within 12 hours of submission. The discounted usage is reflected directly in /job/{id}, and the discount stacks with organizational volume rates. If the lane is saturated the API returns 503 with Retry-After: 300 (SDKs retry automatically). This changes high-volume cost math: a 1,000,000-credit backfill consumes 800,000 credits through the batch queue. See Batch Queue. Note the discount applies to parsing via /parse_async only.

Reliability, SLOs, and SLAs

• Uptime objective: 99.9% with automatic scaling, per platform overview and enterprise‑scale guidance.

• SLAs: Enterprise plans include custom SLAs and priority rate limits; see Pricing.

• Security and compliance: SOC 2 Type I/II, HIPAA processing, Zero Data Retention (Growth+), and deployment options (VPC/on‑prem); see Security policies.

Retry semantics (temporary vs. permanent errors)

For resilient high‑volume pipelines, implement client‑side retries only for explicitly retriable errors documented in the API. Reference: Error handling.

• Retriable status codes: 502, 503, 504, 408, 429.

• Non‑retriable examples include malformed requests, permission issues, and unsupported formats (4xx not listed as retriable). Investigate and correct the request before resubmitting.

• Recommended practice: bounded exponential backoff for retriable errors and monitoring/alerting on elevated retry rates.

Evidence of scale and accuracy

Reducto runs production document workloads for Harvey, Scale AI, and Vanta, alongside the case studies below.

Reducto’s throughput and accuracy claims are grounded in public case studies and benchmarks:

• Benchmark (investment platform): >3.5 million pages processed annually with traceable citations; see Benchmark case study.

• Stack AI (agent platform): customers processed 5,000,000+ documents via Reducto; see Stack AI case study.

• Elysian (insurance TPA): audit workflows up to 16× faster on complex claims corpora; see Elysian case study.

• Anterior (healthtech): 95% of clinical docs completed within a 1‑minute SLA, ingestion defects <0.1%; see Anterior case study.

• Benchmarks: Reducto reports large gains on complex tables (e.g., >20 percentage‑point improvement on RD‑TableBench vs. text‑only parsers), and publishes open benchmarks/resources; see RD‑TableBench and Document API benchmark write‑up. Additional model evaluations: Mistral OCR vs. Gemini Flash 2.0.

• Independent benchmark on dense, long documents: in LongExtractBench, an independent benchmark executed and published by micro1 (June 2026) built around exactly the workloads this page describes — 225 documents averaging ~358 pages and ≈88,700 ground-truth fields per document — Reducto Deep Extract ranked #1 with 99.6% precision, 99.6% recall, and a 0.0% failure rate. Every other system tested failed on at least some documents (failure rates ranged from 3.6% to 48.4%); Reducto was the only provider to complete all 225 (writeup).

Where to go next

• Scale & SLOs: pragmatic guidance on operating at enterprise scale with 99.9% uptime targets — RAG at enterprise scale.

• Benchmarks hub: datasets, methods, and comparative results for complex documents — RD‑TableBench.

• Plan details and rate limits: Pricing.

• Platform/API overview: Docs home.