Read any form with AI vision, no template required
AI-vision extraction is the reading engine inside Voxket OCR. Instead of matching a fixed template, it sends each page of a scanned PDF to a vision LLM (Google Gemini) that reads the page as an image and returns structured fields, text, checkboxes, and signatures, each with a bounding box and a confidence score. No per-layout template to build or maintain.
What this changes for your team
- No template maintenance tax: the model reads the page the way a person would and generalizes across layouts.
- Captures the messy stuff: checkboxes and signatures that template systems miss.
- Faster time-to-value on a new form type: adapt by editing the agent config, not building a template.
- Structured, not raw text: output is machine-parseable JSON, validated against a schema.
From setup to running
- 01
Render
Each PDF page is rasterized to a high-quality image and encoded for efficient, consistent model input.
- 02
Prompt
A configurable OCR agent instructs the model to extract every field as key, value, type, bounding box, and confidence, and to detect checkboxes and signatures.
- 03
Structured output
The model returns JSON validated against a response schema (temperature-zero), so the output shape is consistent.
- 04
Confidence + coordinates
Each field carries a 0–100 confidence score and a bounding box, which power downstream provenance and review.
What's inside
Template-free reading
Works across varied layouts with no per-form template library.
Checkbox detection
Checked vs. unchecked states captured as fields.
Signature detection
Flags signed or initialed regions.
Per-field bounding boxes
The coordinates that make source provenance possible.
Per-field confidence scores
0–100, surfaced in review for triage.
Configurable OCR agent
Prompt and model tunable without a code change.
Built-in error resilience
JSON-repair retries and a transient-failure circuit breaker keep one hiccup from failing a batch.
Where it fits
Common questions
Does it need a template for each form?
No. It reads the page as an image with a vision LLM, so there's no per-layout template to build or maintain.
Can it read checkboxes and signatures?
Yes. Checkboxes are captured as checked/unchecked fields, and signature regions are detected.
What model does it use?
Google Gemini via the vision API, with structured JSON output constrained by a response schema. The model is configurable per form type.
Is the output structured or just text?
Structured. Each field is a key/value with a type, a bounding box, and a confidence score, validated against a schema.
Ready to put an AI workforce on your frontline?
Deploy voice, chat, and video agents that work 24/7, act in your systems, and hand off to your team when it counts. Build your first agent in minutes.
White-glove onboarding for teams · Talk to a human, not a form.