OCR-to-vision routing

Docgate's job is to govern an extraction ladder, not to replace its providers.

For a mixed bundle such as a CV, experience letter, PAN card, and Aadhaar card:

1. Use need and a local selection policy to avoid processing irrelevant pages where local evidence is available.
2. Send selected image-only pages to a lower-cost OCR provider.
3. Apply a local TextQualityGate to the returned transcription.
4. Escalate only weak output to the next provider, normally a vision model.
5. Inspect the receipt: it records selection skips, OCR quality rejections, fallback attempts, cost, and any pages still unknown.

The quality gate is a local minimum-evidence check, not a semantic evaluator. If it rejects a successful OCR response, docgate records quality_gate_rejected as a degradation and tries the next provider. The OCR provider is not treated as unhealthy solely because its text was insufficient.

from decimal import Decimal
import docgate

ocr = docgate.MistralOcrProvider(
    api_key="...",
    # Set this from your own current Mistral pricing agreement.
    cost_per_page=Decimal("0.004"),
)
vision = docgate.OpenAICompatibleProvider(
    model="your-vision-model",
    api_key="...",
    base_url="https://your-host/v1",
    cost_per_page=Decimal("0.01"),
)

result = docgate.read(
    "employment-bundle.pdf",
    need="CV and experience letter",
    selection="top_k",
    max_paid_pages=3,
    providers=[ocr, vision],
    quality_gate=docgate.TextQualityGate(
        min_characters=40,
        min_alphanumeric_characters=25,
        min_words=6,
    ),
    budget=Decimal("0.05"),
)

if not result.receipt.complete:
    # A policy decision belongs to the application, never docgate.
    print(result.receipt.unknown)

The quality gate measures only obvious local evidence, not document meaning. Tune it for the shortest valid document you expect. A short PAN or Aadhaar identifier should use a smaller threshold than a CV or experience letter.