feat(ocr): add Baidu PaddleOCR as a selectable OCR provider

.env.example

@@ -76,7 +76,26 @@ MISTRAL_OCR_MODEL=mistral-ocr-latest
 MISTRAL_OCR_API_URL=https://api.mistral.ai/v1/ocr
 OCR_SEGMENT_PAGES=20
 OCR_REQUEST_TIMEOUT_SECONDS=300
-# Alternatives: OCR_PROVIDER=mock|pypdf|chandra
+# Alternatives: OCR_PROVIDER=mock|pypdf|chandra|paddleocr
+
+# --- OCR (Baidu PaddleOCR / PP-OCR, local CPU) ---
+# OCR_PROVIDER=paddleocr
+# Model tier: medium (34.5M params, default) | small (7.7M) | tiny (1.5M).
+# Or an explicit pair: PADDLEOCR_MODEL=PP-OCRv6_medium_det,PP-OCRv6_medium_rec
+# PADDLEOCR_MODEL=medium
+# PADDLEOCR_OCR_VERSION=PP-OCRv6
+# PADDLEOCR_LANG=
+# PADDLEOCR_DEVICE=cpu
+# PADDLEOCR_DET_MODEL=
+# PADDLEOCR_REC_MODEL=
+# PADDLEOCR_MIN_SCORE=0
+# PADDLEOCR_IMAGE_DPI=192
+# PADDLEOCR_MAX_WORKERS=1
+# Leave false. PaddlePaddle 3.3 + oneDNN crashes on CPU (PIR -> oneDNN conversion);
+# only set true if you are on a build where that crash does not occur.
+# PADDLEOCR_ENABLE_MKLDNN=false
+# Where the checkpoint is cached (set in docker-compose to the mounted volume).
+# PADDLE_PDX_CACHE_HOME=/opt/hf-cache/paddlex
 
 # --- LLM: Gemma 4 (translation + domain tagging) ---
 AGRINET_GEMMA_BASE_URL=https://dev-vistaar.da.gov.in/gemma4

Dockerfile

@@ -4,11 +4,16 @@ FROM python:3.10-slim
 WORKDIR /app
 
 # Install system dependencies
-# apt-get is skipped — curl is not needed at runtime.
-# Uncomment below only if you need libreoffice (office→PDF) or curl:
-# RUN apt-get update && apt-get install -y --no-install-recommends \
-#     curl \
-#     && rm -rf /var/lib/apt/lists/*
+# libgomp1  — OpenMP runtime; paddlepaddle's core library links against it and
+#              refuses to import without it ("Can not import paddle core ...").
+# libgl1 + libglib2.0-0 — the runtime shared libraries opencv-contrib-python
+#              (a PaddleX dependency) loads at import time.
+# curl is not needed at runtime; add it only if you add a healthcheck that uses it.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libgomp1 \
+    libgl1 \
+    libglib2.0-0 \
+    && rm -rf /var/lib/apt/lists/*
 
 COPY --from=ghcr.io/astral-sh/uv:0.12.5 /uv /uvx /usr/local/bin/
 

ENV.md

@@ -111,7 +111,7 @@ OCR is a **separate process** (not in `docker-compose`). The worker calls
 
 | Variable | Default | Purpose |
 |----------|---------|---------|
-| `OCR_PROVIDER` | `chandra` | OCR provider (`chandra`, `mock`, `pypdf`) |
+| `OCR_PROVIDER` | `chandra` | OCR provider (`chandra`, `mistral`, `paddleocr`, `mock`, `pypdf`) |
 | `OCR_MODEL` | `chandra` | Model name |
 | `CHANDRA_VLLM_BASE_URL` | `''` | Base URL; local HF default `http://localhost:8010/v1` |
 | `CHANDRA_OCR_API_URL` | `''` | Alternate full OCR API URL (overrides base path build) |
@@ -124,6 +124,29 @@ OCR is a **separate process** (not in `docker-compose`). The worker calls
 | `OCR_SEGMENT_PAGES` | `20` | Segment size for long docs |
 | `CHANDRA_HF_HOME` / `HF_HOME` | — | Hugging Face cache (HF server / scripts) |
 
+**Baidu PaddleOCR (local CPU lane):**
+
+`OCR_PROVIDER=paddleocr` runs PP-OCR in-process — no GPU, no remote inference server.
+The checkpoint is downloaded on first use into `PADDLE_PDX_CACHE_HOME` (set by
+`docker-compose.yml` to a directory on the mounted `HF_HOME` volume, so it survives a
+rebuild). Available on the **linux/amd64** image only: `paddlepaddle` publishes no
+aarch64 wheels.
+
+| Variable | Default | Purpose |
+|----------|---------|---------|
+| `PADDLEOCR_MODEL` | `medium` | PP-OCRv6 tier (`medium` \| `small` \| `tiny`), or an explicit `"<det>,<rec>"` model-name pair |
+| `PADDLEOCR_OCR_VERSION` | `PP-OCRv6` | `PP-OCRv6` \| `PP-OCRv5` \| `PP-OCRv4` |
+| `PADDLEOCR_LANG` | `''` | Checkpoint language hint (`en`, `ch`, `dev`, …); empty = the checkpoint default |
+| `PADDLEOCR_DET_MODEL` | `''` | Explicit detector model name (wins over `PADDLEOCR_MODEL`) |
+| `PADDLEOCR_REC_MODEL` | `''` | Explicit recogniser model name (wins over `PADDLEOCR_MODEL`) |
+| `PADDLEOCR_DEVICE` | `cpu` | `cpu` or `gpu` / `gpu:0` |
+| `PADDLEOCR_MIN_SCORE` | `0` | Drop recognised lines below this confidence |
+| `PADDLEOCR_IMAGE_DPI` | `192` | Page render DPI |
+| `PADDLEOCR_MAX_WORKERS` | `1` | Parallel OCR workers |
+| `PADDLEOCR_ENABLE_MKLDNN` | `false` | Intel oneDNN acceleration. Leave `false`: PaddlePaddle 3.3 crashes on CPU in the PIR → oneDNN conversion (`ConvertPirAttribute2RuntimeAttribute not support`) and the `FLAGS_*` env vars do not work around it |
+| `PADDLE_PDX_CACHE_HOME` | — | Where the checkpoint is cached (`/opt/hf-cache/paddlex` in compose) |
+| `PADDLE_PDX_MODEL_SOURCE` | `huggingface` | PaddleX model host; `modelscope` is the alternative |
+
 **Local start (real model, GPU + torch + model download):**
 
 ```bash

docker-compose.yml

@@ -254,6 +254,9 @@ services:
       # HF cache path propagated into the container
       HF_HOME: ${HF_HOME:-/opt/hf-cache}
       HF_HUB_CACHE: ${HF_HOME:-/opt/hf-cache}
+      # PaddleX (Baidu PaddleOCR) weight cache — kept on the same mounted volume so
+      # the on-demand checkpoint download survives a container rebuild.
+      PADDLE_PDX_CACHE_HOME: ${HF_HOME:-/opt/hf-cache}/paddlex
     ports:
       - "127.0.0.1:${API_HOST_PORT:-8011}:8001"
     depends_on:
@@ -295,6 +298,9 @@ services:
       ALLOWED_FILE_PATHS: /app/books,/data/documents
       HF_HOME: ${HF_HOME:-/opt/hf-cache}
       HF_HUB_CACHE: ${HF_HOME:-/opt/hf-cache}
+      # PaddleX (Baidu PaddleOCR) weight cache — the worker is the process that
+      # actually runs OCR, so the checkpoint is downloaded and cached here.
+      PADDLE_PDX_CACHE_HOME: ${HF_HOME:-/opt/hf-cache}/paddlex
     depends_on:
       temporal:
         condition: service_healthy

pipeline/ocr/__init__.py

@@ -4,6 +4,7 @@
 from .chandra_vllm import ChandraVllmOcrProvider
 from .mistral_ocr import MistralOcrProvider
 from .mock import MockOcrProvider
+from .paddleocr_local import PaddleOcrProvider
 from .service import get_ocr_provider, load_ocr_config, ocr_pdf, ocr_pdf_in_segments
 from .text_layer import split_pdf_range
 
@@ -13,6 +14,7 @@
     "MockOcrProvider",
     "OcrConfig",
     "OcrProvider",
+    "PaddleOcrProvider",
     "get_ocr_provider",
     "load_ocr_config",
     "ocr_pdf",

pipeline/ocr/base.py

@@ -32,6 +32,19 @@ class OcrConfig:
     image_dpi: int = 192
     request_timeout_seconds: float = 300.0
 
+    # Baidu PaddleOCR (local, CPU) lane. These are separate fields rather than
+    # reusing ``model``/``image_dpi`` because the PaddleOCR knobs are spelled
+    # differently from the Chandra/Mistral ones, and one deployment may want both
+    # lanes configured at once (switch with OCR_PROVIDER).
+    paddleocr_version: str = "PP-OCRv6"
+    paddleocr_lang: str = ""
+    paddleocr_device: str = "cpu"
+    paddleocr_det_model: str = ""
+    paddleocr_rec_model: str = ""
+    paddleocr_min_score: float = 0.0
+    paddleocr_image_dpi: int = 192
+    paddleocr_enable_mkldnn: bool = False
+
 
 PageDict = dict