---
title: Chandra OCR 2
emoji: 📄
colorFrom: indigo
colorTo: purple
sdk: gradio
app_file: app.py
pinned: false
license: apache-2.0
short_description: Layout-aware document OCR to markdown + extracted figures
models:
  - datalab-to/chandra-ocr-2
tags:
  - ocr
  - document-ai
  - vision-language
suggested_hardware: zero-a10g
---

# Chandra OCR 2 — Space demo

Gradio demo for [`datalab-to/chandra-ocr-2`](https://huggingface.co/datalab-to/chandra-ocr-2),
Datalab's layout-aware document OCR model. Upload PDFs or images and get back
markdown with tables, math, forms, and reading order preserved — **plus** a
separate figure-extraction stage that crops charts, diagrams, photos and stamps
at full resolution and reports the model's caption and any structured data it
read from them.

## Hardware

**This will not run on the free CPU tier.** The model is ~10B parameters,
roughly 20 GB in bf16.

| Hardware | Works? |
|---|---|
| CPU basic (free) | No — OOM at load |
| ZeroGPU (H200 slice) | Yes — recommended |
| RTX PRO 6000 (Blackwell, sm_120) | Yes |
| L40S / A100 | Yes |
| T4 (16 GB) | No, unless you add 4-bit quantisation |

Set this under **Settings → Hardware** after creating the Space.
`suggested_hardware` in the frontmatter is only a hint to visitors; it does not
provision anything.

## How it works

- PDFs are rasterised page by page with PyMuPDF (no poppler/apt needed).
- Each page is capped on its long edge before inference — visual token count
  scales with area, so this is the main lever on latency and memory.
- Pages are batched `PAGES_PER_GPU_CALL` at a time so each ZeroGPU allocation
  finishes inside its duration budget.
- **Stage 1 (text):** raw layout output → markdown via `parse_markdown`.
- **Stage 2 (figures):** `parse_chunks` locates figure-like blocks, which are
  cropped from the **full-resolution** source page (the bbox is per-axis
  normalised, so it maps cleanly onto any resolution with the same aspect
  ratio). Captions and structured data come from the model's layout output and
  are labelled **degraded** when the fallback path was used.

Output is offered as rendered markdown, markdown source, the raw model string,
a figure gallery + metadata table, a combined `.md`, and a `.zip` containing
per-page markdown, per-page raw output, extracted figure images, and a
`manifest.json`.

The app prefers the official `chandra` package (`generate_hf` + `BatchInputItem`
+ `parse_markdown` + `parse_chunks`). If that import fails, it falls back to
driving the chat template through plain `transformers` so the Space still boots
— and reports the degradation in the status panel and diagnostics.

## Configuration

Edit the constants at the top of `app.py`:

| Constant | Default | Purpose |
|---|---|---|
| `PAGES_PER_GPU_CALL` | 3 | Pages per ZeroGPU allocation |
| `GPU_DURATION` | 180 | Seconds requested per allocation |
| `MAX_PAGES` | 20 | Per-run page cap |

Prompt types are **enumerated from the package** (`PROMPT_MAPPING`), not
hardcoded. Only `ocr_layout` and `ocr` are currently exposed; the dropdown is
editable so you can try undocumented values.

## Troubleshooting

| Symptom | Cause / fix |
|---|---|
| Space won't start, opaque "upload failed" | Model load failed at import — the diagnostics panel auto-opens with the exact error. |
| `CUDA out of memory` per page | Lower **Max image side**; it's the biggest VRAM lever. |
| No figures in the gallery | The model returned no figure bboxes for this prompt type, or the `chandra` package is unavailable (outputs marked degraded). |
| `flash-attn` build error | Do **not** add flash-attn — Blackwell (sm_120) has no prebuilt wheels. SDPA is used. |
| `torch` version mismatch | Do **not** pin torch — the Space image ships a build matched to the driver/CUDA. |

## Licence — read before making this public

The Space **code** here is Apache-2.0. The **model weights** are not:

> Code is Apache 2.0. Model weights are under a modified OpenRAIL-M license.
> Free for research, personal use, and startups under $2M funding/revenue.
> Cannot be used competitively with our API.

A public, free, hosted OCR endpoint is plausibly "competitive with our API."
If you are past the revenue threshold, or intend this as a product rather than a
demo, check with Datalab first. Setting the Space to **private** avoids the
question entirely.