Fast, CPU-first OCR changes how large document collections are processed: by trading a slightly smaller accuracy ceiling for vastly lower cost and latency, Indra OCR makes local and batch document parsing practical without GPUs. It bundles a 1M-parameter recognizer in ~18 MB of ONNX models and a pipeline that returns lines, layout blocks, reading order and markdown or JSON outputs.
What Sets It Apart
- High CPU throughput and low latency — engineered to run one single-threaded worker per core, typical runs hit ~7.3 pages/s on 4 cores and scale to ~49 pages/s on 32 cores, with a median ~230 ms per page on a 4-core x86 run. So what: large archives and automated pipelines can be processed much cheaper and faster than GPU or heavy-Python alternatives.
- Very small, dependency-light models — ONNX runtime-only models (~18 MB total, no PyTorch) and a ~1M-parameter recognizer. So what: easy deployment in constrained environments, fast cold-starts and simple CI/edge packaging.
- Layout-aware outputs and hybrid PDF handling — emits per-line boxes, layout blocks with labels (headings, lists, tables, captions, footnotes, headers), reading order, and uses a PDF text-layer + OCR hybrid mode. So what: downstream tasks (indexing, extraction, conversion to Markdown/HTML) get structured outputs, reducing post-processing work.
- Tunable math support and accuracy trade-offs — optional math recognizer boosts headline/text scores (math mode raises headline and text metrics) at a cost in speed. So what: you can enable math when accuracy on formulas matters and disable it for maximum throughput.
Who It’s For and Trade-offs
Great fit if you need high-throughput, on-prem or edge document OCR with layout awareness — batch PDF pipelines, local privacy-preserving processing, or systems where GPUs are unavailable or too costly. It is also convenient when you want small deployable artifacts (ONNX) and a CLI + Python API for integration.
Look elsewhere if your primary needs are high-accuracy handwriting recognition, heavily degraded or low-resolution scans, vertical text, or very complex tables with merged/nested headers — Indra favors speed and compactness over solving those niche, high-difficulty cases. For maximum accuracy on handwriting, complex tables or degraded scans, use a more heavyweight managed model or specialized OCR service.
