Batch Processing

Process multiple documents efficiently.

---

Quick Start

from pathlib import Path
from omnidocs import Document
from omnidocs.tasks.text_extraction import QwenTextExtractor
from omnidocs.tasks.text_extraction.qwen import QwenPyTorchConfig

# Initialize once (expensive)
extractor = QwenTextExtractor(backend=QwenPyTorchConfig(device="cuda"))

# Process all PDFs
for pdf_path in Path("documents/").glob("*.pdf"):
    doc = Document.from_pdf(pdf_path)

    for i, page in enumerate(doc.iter_pages()):
        result = extractor.extract(page, output_format="markdown")

        output = Path("output") / f"{pdf_path.stem}_page_{i+1}.md"
        output.write_text(result.content)

---

With Progress Tracking

import time
from pathlib import Path

pdf_files = list(Path("documents/").glob("*.pdf"))
start = time.time()

for idx, pdf_path in enumerate(pdf_files, 1):
    doc = Document.from_pdf(pdf_path)

    for page in doc.iter_pages():
        result = extractor.extract(page)

    # Progress
    elapsed = time.time() - start
    remaining = (len(pdf_files) - idx) * (elapsed / idx)
    print(f"[{idx}/{len(pdf_files)}] {pdf_path.name} - ETA: {remaining/60:.1f}min")

---

Memory Management

Clear cache periodically for large batches:

for i, page in enumerate(doc.iter_pages()):
    result = extractor.extract(page)
    save_result(result)

    # Free memory every 10 pages
    if i % 10 == 0:
        doc.clear_cache()

---

Stream to Disk

Don't accumulate results in memory:

import json

with open("results.jsonl", "w") as f:
    for pdf_path in pdf_files:
        doc = Document.from_pdf(pdf_path)
        result = extractor.extract(doc.get_page(0))

        record = {"path": str(pdf_path), "word_count": result.word_count}
        f.write(json.dumps(record) + "\n")

---

Error Handling

results = []
errors = []

for pdf_path in pdf_files:
    try:
        doc = Document.from_pdf(pdf_path)
        result = extractor.extract(doc.get_page(0))
        results.append({"path": str(pdf_path), "success": True})
    except Exception as e:
        errors.append({"path": str(pdf_path), "error": str(e)})

print(f"Succeeded: {len(results)}, Failed: {len(errors)}")

---

Performance Tips

Tip	Why
Initialize extractor once	Model loading takes 2-3s
Use VLLM for large batches	2-4x better throughput
Stream results to disk	Constant memory usage
Clear cache periodically	Prevents OOM

---

For Cloud Scale

See Deployment for processing on Modal GPUs.