LightOnOCR-3-4B

# LightOnOCR-3-4B runs on your side, as in the model card: render PDF pages
# at 400 DPI with a 2048 px longest edge, then call the model with the
# "grounding" prompt. No Mixpeek extractor runs these weights.
import re, requests, torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "lightonai/LightOnOCR-3-4B"
model = Qwen3_5ForConditionalGeneration.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)

conversation = [{"role": "user", "content": [
    {"type": "image", "url": "https://your-cdn.example.com/reports/2026-q3/page-004.png"},
    {"type": "text", "text": "grounding"},
]}]
inputs = processor.apply_chat_template(
    conversation, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt", enable_thinking=False,
).to(model.device)
output_ids = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.2)
page = processor.decode(output_ids[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True)

# Each block starts with ![label](x0,y0,x1,y1). Upload one object per block to
# a bucket whose collection uses features: ["text_search"] (text_extractor,
# E5-Large). The text extractor reads up to 512 tokens, so block-sized
# objects keep long pages from being truncated. List source_key, page, label
# and bbox in the collection's field_passthrough so they reach each document.
MARK = re.compile(r"!\[([a-z_]+\+?)\]\((\d+),(\d+),(\d+),(\d+)\)\n")
parts = MARK.split(page)
for i in range(1, len(parts), 6):
    label, x0, y0, x1, y1, text = parts[i:i + 6]
    if label.rstrip("+") in ("header", "footer", "page_number"):
        continue
    requests.post(
        "https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
        headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
        json={
            "key_prefix": "/reports/2026-q3",
            "blobs": [{"property": "block_text", "type": "text", "data": text.strip()}],
            "source_key": "reports/2026-q3.pdf", "page": 4, "label": label,
            "bbox": [int(x0), int(y0), int(x1), int(y1)],
        },
    )