---
type: "post"
id: "indra-ocr-1-1-1-and-1-1-63fe4c"
slug: "indra-ocr-1-1-1-and-1-1-63fe4c"
title: "Indra OCR 1.1.1 and 1.1.2 fix visual-order RTL PDFs and legacy-font Hindi text layers"
description: "Datalab’s Indra OCR 1.1.1 and 1.1.2 change when the tool trusts text already embedded in a PDF: it now OCRs pages when that layer misrepresents right-to-left text or legacy-font Hindi."
retrieval_nugget: "Datalab’s Indra OCR 1.1.1 and 1.1.2 change when the tool trusts text already embedded in a PDF: it now OCRs pages when that layer misrepresents right-to-left text or legacy-font Hindi. The changelog reports substantial gains on affected test pages, addressing a source of silent corruption in document pipelines."
published_at: "2026-11-11"
updated_at: "2026-11-11"
record_date: "2026-11-11"
date_kind: "published_at"
topics: ["multimodal","tools"]
entities: ["Datalab","Indra OCR","Kruti Dev","DevLys","Bijoy"]
related_patterns: []
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://github.com/datalab-to/indra/blob/main/CHANGELOG.md"
source_urls: ["https://github.com/datalab-to/indra/blob/main/CHANGELOG.md"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:indra-ocr-1-1-1-and-1-1-63fe4c"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/indra-ocr-1-1-1-and-1-1-63fe4c/","markdown":"https://newruntime.com/posts/indra-ocr-1-1-1-and-1-1-63fe4c.md","json":"https://newruntime.com/posts/indra-ocr-1-1-1-and-1-1-63fe4c.json"}
markdown_url: "https://newruntime.com/posts/indra-ocr-1-1-1-and-1-1-63fe4c.md"
---

# Indra OCR 1.1.1 and 1.1.2 fix visual-order RTL PDFs and legacy-font Hindi text layers

## Retrieval answer

Datalab’s Indra OCR 1.1.1 and 1.1.2 change when the tool trusts text already embedded in a PDF: it now OCRs pages when that layer misrepresents right-to-left text or legacy-font Hindi. The changelog reports substantial gains on affected test pages, addressing a source of silent corruption in document pipelines.

## Summary

Datalab’s Indra OCR 1.1.1 and 1.1.2 change when the tool trusts text already embedded in a PDF: it now OCRs pages when that layer misrepresents right-to-left text or legacy-font Hindi. The changelog reports substantial gains on affected test pages, addressing a source of silent corruption in document pipelines.

## What happened

Datalab’s Indra OCR changelog lists two point releases, 1.1.1 and 1.1.2. Both use the same models as Hugging Face tag v1.1.0; the changes described are to how Indra handles PDF text layers. A PDF can contain both a rendered page image and an embedded text layer. Indra can use that layer rather than recognizing the image, but the changelog describes cases where the layer’s characters or ordering do not faithfully represent what a reader sees.

Version 1.1.1 addresses right-to-left text layers stored partly in visual order. The changelog says some PDF producers, including Chrome, store Arabic-script letters this way. Read as ordinary text, these layers can reverse letters and insert spaces within words. Indra now declines to trust such a layer and OCRs the page image instead. On Chrome-printed Wikipedia pages, reported word recall rises from 0.17 to 1.00 for Arabic, 0.25 to 0.89 for Persian, 0.22 to 0.95 for Pashto, and 0.28 to 0.98 for Urdu. The release also prevents a right-to-left page from being mistaken for vertical writing. The changelog says Hebrew and Yiddish layers in logical order continue to be read as text.

Version 1.1.2 targets legacy-font text layers in Hindi PDFs. The changelog names Kruti Dev and DevLys: these fonts can encode Devanagari text as Latin codes in the PDF layer, so trusting the layer produces Latin-looking garbage rather than Hindi. Before relying on a text layer, Indra now compares the script of a few of its lines with the corresponding line images. If the layer appears Latin while the page image shows another script, Indra runs OCR on the page. For the affected Hindi pages, reported word accuracy rises from 0.01–0.20 to 0.85–0.89. The same check covers other legacy fonts, including Bijoy for Bangla. Datalab says the check adds about 5 milliseconds per text-layer page.

The changelog presents these as targeted detection-and-fallback changes, not as new model releases: both versions retain the same models. Its reported gains are specific to the affected document types and test examples, rather than a claim of improved accuracy on every PDF or language.

## What the archive adds

New Runtime’s earlier coverage of [Indra OCR 1.1.0](https://newruntime.com/posts/release-indra-ocr-1-1-0-datalab-to-ac3594/) described handwriting recognition, support for more than 80 languages on scans, and rotated-page correction while remaining CPU-runnable. These two point releases add new handling for broken PDF text layers to that known product story: they are new fixes, not a restatement of the 1.1.0 release.

## Trend signal

This update does not yet fit a pattern New Runtime tracks. It improves a specific document-processing behavior, but the changelog does not establish a broader shift in agent architectures, protocols, or deployment economics.

## Why it matters

PDF pipelines often encounter documents that look correct to a person but contain a defective text layer. If software trusts that layer, downstream systems may receive reversed words, misplaced spaces, or the wrong writing system without an obvious processing error. A targeted comparison between embedded text and page imagery gives Indra a way to fall back to OCR for these cases while retaining the cheaper text-layer route where it works.

For teams building search, extraction, or document-processing workflows, the update is relevant when inputs include PDFs created by varied software or legacy-font documents. The reported results suggest that a small amount of detection can matter more than blindly choosing either embedded text or full-page OCR. The changelog does not establish that the method generalizes to every producer, font, or script, so teams should validate against their own document samples before treating the fallback as a complete fix.

## Limits and uncertainty

The changelog reports results for Chrome-printed Wikipedia pages and affected Hindi pages, but it does not establish performance across all PDF producers, document collections, or fonts. The listed recall and accuracy gains are not accompanied here by sample sizes or a description of the evaluation procedure. Datalab reports an approximately 5 ms cost per text-layer page for the Hindi check; the changelog does not quantify the cost of the right-to-left fallback or the frequency of either failure mode in production. The release notes also do not establish whether every visually and textually mismatched layer will be detected.

## Skill to build now

Take a small sample of PDFs from your actual ingestion pipeline and compare extracted text with the visible page, especially for right-to-left languages and older Hindi or Bangla documents. Record whether the embedded layer is present, whether its reading order and script match the image, and what Indra returns. Include a few known-good documents as controls so a fallback does not appear successful merely because the test set contains only failures. This gives a team a practical regression set for checking whether the new behavior addresses its own inputs.

## Reader takeaway

Indra OCR’s 1.1.1 and 1.1.2 releases make PDF text extraction more defensive: when the embedded layer’s script or ordering conflicts with the rendered page, the tool can recognize the image instead. The gains are promising for affected multilingual documents, but teams should test representative PDFs rather than assume the fixes cover every malformed text layer.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- topic hub: [Tools](https://newruntime.com/topics/tools.md): Explore the tools topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)