Extract Text (vLLM)

How it fits

Typical backends

Caveats

• CompatibilityA local vLLM sidecar serving the chosen OCR model must be reachable; the model is fetched and loaded by that sidecar, so the first deployment of an uncached checkpoint pays the download and warm-up cost before serving.
• I/O contractThe output arm is fixed when the backend deploys: a text consumer pins the String transcript, a region consumer pins per-block records with Polygon<Double> or Rectangle<Double> bounds. The bounded arms are filled from the model's layout JSON; a model that returns prose only fills a single page-spanning block.
• Parameter interactionprompt and model_name must agree — the default prompt is the dots.ocr layout-parsing instruction and only a layout-emitting model produces the per-block JSON the bounded arms read; a general vision model needs a plain extraction prompt and supports the String arm only.
• Fallbackdots.ocr does not return per-block scores, so the confidence arms report a fixed full score for every detected block rather than a model-derived probability.
• State lifetimetemperature, max_tokens and timeout_seconds are re-read per page, while model_name, prompt, image_format and jpeg_quality are bound once at startup.
• CostEach page is re-encoded to image_format.jpeg or image_format.png and base64-embedded into the request, so high resolutions and dense pages inflate payload size and per-call latency.