# By combining Kognitos and GPT3, Optical Character Recognition is taking a huge leap forward

> See how Kognitos teaches OCR location-based logic in English — recovering a customer ID from below a trailer number and storing the technique as a reusable learning.

**Page**: https://www.kognitos.com/videos/by-combining-kognitos-and-gpt3-optical-character-recognition-is-taking-a-huge-leap-forward/
**Watch on YouTube**: https://www.youtube.com/watch?v=1-e0xPbp1y8
**Length**: 3m 16s

## What's in this video

A 3-minute demonstration of how Kognitos combines traditional OCR with English-defined logic to extract fields that template-based OCR cannot — and how every fix becomes a permanent learning.

## The OCR problem this video solves

Most OCR pipelines need a hand-built template per document layout. Damaged, unstructured, or unfamiliar documents break them. A human reading the document would say something like “the customer ID is always the line below the trailer number” — simple logic, but until now hard for automation to use.

## How Kognitos teaches OCR new tricks

- **Base OCR pass**: Kognitos runs traditional OCR first and reports a confidence score for every field, alongside the values it could and could not extract.
- **The exception**: in this run, the customer ID could not be extracted, so the brain pauses and surfaces the gap.
- **Mini playground**: a sandboxed test environment where the user types phrases like “grab the line below the document's trailer number” and verifies the result without touching the real system of record.
- **Validate and save**: once the value comes out correct, the user clicks *Teach a Technique* and the rule is saved.
- **Fallback at scale**: every future document from that vendor runs regular OCR first; if the customer ID is still missing, Kognitos applies the saved location-based rule automatically.
- **Manage techniques**: all learnings for a process live in one place — they can be reviewed, edited, or removed.

## Why pair OCR with an LLM brain

OCR alone can read pixels. An LLM alone hallucinates. Kognitos combines them: OCR does the deterministic reading, the LLM-driven brain understands the English logic that fills the gaps, and a learnings library makes the combination repeatable on damaged or variable documents.

## FAQs

**Q: What's the example used in this OCR demo?**

A vendor document where regular OCR successfully extracts most fields but cannot find the customer ID. The user teaches Kognitos a location-based rule — “the customer ID is the line below the trailer number” — to handle that vendor going forward.


**Q: What is the Kognitos mini playground used for?**

It's a sandboxed test environment where you can try different English phrases or location rules against a document and see the result, without touching any live application. Once the logic works, you save it as a technique.


**Q: How does Kognitos use the saved technique on future documents?**

For each new document from the same vendor, Kognitos runs traditional OCR first. If the field it needs is missing or low-confidence, it falls back to the saved location-based rule — recovering the value without human intervention.


**Q: Can I review and remove techniques later?**

Yes. All learnings for a process are stored in a managed list where you can review, edit, or delete them at any time.