Choose theme
AI Resources
Indra OCR
Indra OCR is Datalab's CPU-based toolkit for turning PDFs and images into Markdown, plain text or structured JSON. It combines text recognition with layout blocks and reading order.
It runs through ONNX Runtime without a GPU or PyTorch. A command-line tool, Python interface and local web viewer offer different ways to extract text and examine the result. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
Document text and structure on a CPU
OCR turns text in a page image into machine-readable characters. Indra accepts PDFs and common image formats, then returns text alongside line positions and labelled page blocks such as headings, lists, captions and tables.
Why it stands out
Existing PDF text can skip recognition
In its default automatic mode, Indra reads a PDF's text layer when it passes the project's checks and applies OCR where needed. A page can therefore combine existing text with recognized text inside images, rather than treating every page as a fresh scan.
Availability
Python package, source and model files
The indra-ocr package, repository and model files are public. The project declares Apache-2.0 for the code and a modified OpenRAIL-M licence for the weights; see the main repository and model card for current terms.
Why it matters
What makes it useful
For a folder of printed reports or scanned reference pages, Indra provides a route to text that a search or document-processing workflow can consume. Batch conversion accepts multiple inputs and can use several CPU cores. Markdown keeps headings and lists, while JSON includes line boxes, block labels and reading order when the next step needs more than a flat transcript.
What to know
Where it fits
Indra sits at the extraction step before search, retrieval or summarisation. Install the Python package, supply a PDF or image, and choose the output format. The viewer overlays detected lines and blocks on the page and exposes the extracted text, confidences and timings. It gives a way to compare the output with the page before passing that text into another system.
Notable points
What stands out
The headline "1M parameters" describes the small main recognizer, not every component in the default pipeline. The README gives about 1.25 million parameters for that recognizer and 7.3 million across the default models, with a roughly 76 MB download including language models. Optional math recognition adds separate models and writes formulas as LaTeX; it is off by default and trades additional processing time for that output.
Before using
What to review
The documented setup needs Python 3.10 or newer. The project lists testing on Linux and Apple-silicon macOS; that is not a native Windows testing claim. Models download from Hugging Face on first use, with a local model-directory option also documented.
Datalab describes Indra as built for clean printed pages. Its limitations include handwriting, degraded scans, complex tables, some scripts and diacritics, and certain reading-order layouts. A table can have text in the wrong row or column even when individual words look plausible.
The repository's speed and accuracy comparisons are project-reported benchmark results with specified hardware, settings and page sets. They do not establish the result on your documents; a small sample containing your actual layouts and scripts is a useful starting comparison.
Reader fit
Who may find it relevant
Practical builders and technically comfortable document users who want CPU-based extraction are the clearest audience. The CLI suits file conversion, and the Python interface suits a larger pipeline. If your collection mainly contains cursive notes, damaged scans or intricate tables, the project's own limitations make those documents a poor basis for expecting clean output.
Editorial note
Why LifeHubber lists it
LifeHubber lists Indra because its page boxes, reading order and viewer help document-search builders trace extracted text back to the page. That makes it easier to check whether a suspicious passage was misread during extraction or misinterpreted by a later AI system.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
More in Ecosystem
Keep browsing this category
Explore more AI ecosystem resources.
Laya
NandhaKishorM/laya
An early Apache-2.0 model family and Python runtime for bounded choice, score, and yes-or-no decisions, with English, multilingual, and task-specialized checkpoints plus a router that selects between them.
E2B
e2b-dev/E2B
Linux sandboxes for AI-generated code, with Python and JavaScript SDKs, a Code Interpreter, templates and pause/resume workflows through a managed service or public runtime.
OpenSandbox
opensandbox-group/OpenSandbox
General-purpose sandbox runtime infrastructure for AI applications, with multi-language SDKs, unified sandbox APIs, Docker and Kubernetes runtimes, CLI and MCP paths, code-interpreter support, browser and desktop examples, and network-control features for agent workloads.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
