Numbers measured on a hold-out evaluation set of 50,000+ real documents — not curated demos. Every category is tested independently with ground-truth annotations.
Results are reproducible via the public evaluation endpoint. Any team can submit their own document set and receive field-level accuracy reports with confidence scores and bounding-box references for every extracted value.
We measure accuracy at the field level, not the page level. A missed value counts — even when the rest of the document is perfect.
Three AI models run in parallel and vote on each field. Disagreement triggers a fourth pass with the highest-confidence model taking precedence.
Sum checks, date range rules, and format constraints catch AI hallucinations before output reaches your pipeline.
Models are retrained monthly on real-world documents across every supported category. Accuracy only goes up over time.
Single-document latency and batch throughput measured under production load, not lab conditions. Numbers are p50 / p95 / p99 percentiles.
All measurements taken in our US-East production region under standard load, not a dedicated benchmark environment.
Each document type was measured over 1,000 independent requests across a 30-day rolling window.
Latency includes full round-trip: upload, OCR, extraction, validation, and JSON serialization.
Any customer can reproduce these results using the /benchmark endpoint in the API. Results are returned with signed timestamps.
Every microsecond is accounted for. Here's exactly what happens inside the platform from the moment a document hits the API to the moment validated JSON reaches your application.
REST API, direct upload, URL, or cloud storage. Accepts 50+ file types.
Identifies orientation, language, document type, and structure in one pass.
Vision-language model + LLM run in parallel. Each field gets a confidence score.
Sum checks, date ranges, format rules, and schema constraints catch hallucinations.
Schema-constrained JSON ready for any database, ERP, or downstream AI pipeline.
Sync REST response or async webhook. Streaming NDJSON for large batch jobs.
Every step runs in an isolated sandbox. Documents are encrypted in transit (TLS 1.3) and at rest (AES-256). Zero data retention by default — files are purged from memory immediately after extraction completes.
Six dimensions. Three platforms. Scored on publicly available data, independent benchmarks, and verified user reviews.
Every number on this page comes from a reproducible methodology. This section explains what we measured, how we measured it, and how you can verify it yourself.
Most document AI vendors publish accuracy numbers based on cherry-picked document types or internal evaluation sets that don't reflect production conditions. The DocsFlow AI benchmark is designed to be reproducible by any team: submit your own documents via the public evaluation endpoint and receive field-level accuracy reports with confidence scores and bounding-box references for every extracted value.
Accuracy is measured per-field across the evaluation corpus. A document with 20 fields where one is incorrect scores 95% — not 100%. The 99.2% overall figure is the weighted average across all document types and all fields in the 50,000-document evaluation set. Fields with a confidence score below 0.85 are flagged for human review and excluded from the accuracy calculation to prevent low-confidence guesses from inflating the number.
Latency is measured from the moment the API receives the first byte of the uploaded file to the moment the final validated JSON is returned. This includes OCR, extraction, validation, and JSON serialization — the full round-trip. Numbers are reported as p50, p95, and p99 over 1,000 requests per document type in the US-East production region under standard load. We do not run benchmarks in a dedicated low-latency environment.
Base64.ai claims up to 99.7% accuracy on specific document types under controlled conditions. DocsFlow AI reports 99.2% as a weighted average across all document types in its evaluation corpus. The two numbers are not directly comparable because they use different evaluation methodologies. DocsFlow AI's numbers are from a public hold-out set that any team can reproduce; Base64.ai's are from an internal benchmark. For a fair comparison, both platforms should be tested on the same document set.
DocsFlow AI processes batch jobs in parallel rather than sequentially. A batch of 1,000 documents does not take 1,000× the single-document latency — it takes roughly the same time as the longest individual document in the batch, because all documents are processed simultaneously. This architecture is what makes the 10,000+ documents per minute throughput figure possible.
How we measure, what the numbers mean, and how to reproduce them.
