Unstract : Extract Data from Unstructured Docs

Author: Data Science in your pocket · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Unstract : Extract Data from Unstructured Docs · Channel: Data Science in your pocket · 05:50 · Watch on YouTube

Topics: Enterprise AI applications, Data, databases and embeddings, RAG, search and information retrieval

Summary

The video introduces Unstract as a tool for extracting structured data from unstructured documents such as PDFs, images, scans, and handwritten forms. The presenter demonstrates asking questions in natural language and receiving answers or structured JSON, including customer details and applicant information. The platform offers configurable LLMs and other components, and the presenter highlights its model-agnostic setup. The video also shows how to deploy a workflow as an API or ETL pipeline.

Key points

Template-based OCR can fail when document layouts change.
Unstract uses LLMs to extract structured JSON from PDFs, images, and scans.
Users describe the information they want to extract in natural language.
The demo extracts a credit card issuer, customer name, and customer address.
The platform supports changing the LLM and configuring other components.
The presenter tests the tool on a partly handwritten form.
Responses can be constrained to a specified format, including JSON and predefined answer choices.
Unstract can be deployed as an API or used as an ETL pipeline.

Chapters

Start	Chapter	Summary
00:00	Unstructured document extraction and Unstract	The presenter describes the challenges of extracting information from documents and introduces Unstract as an LLM-based tool.
01:34	Extracting information from a credit card document	The demo uses natural-language questions to retrieve the issuer, customer name, and a structured customer address.
03:07	Testing handwritten documents and structured answers	The presenter tests extraction on a partly handwritten form and requests applicant details in specific formats.
04:45	Deploying a workflow as an API	The presenter configures an API deployment and shows its workflow endpoint and code snippet.

Quotes

You define what you want to extract in natural language and boom you get your results within seconds — 00:31

there’s a structured output that we have got — 02:05

it performs well on handwritten documents also — 03:07

Now you can use it as an API or ETL pipeline too. — 05:17

Concepts and entities

Unstract, Unstract Cloud, Prompt Studio, OCR, LLM, JSON, AGPL, ETL pipeline, American Express

Full transcript (original language)

00:00 So hi everyone, welcome back to data science in your pocket and today we would be exploring a very crucial tool that would be helping you in your enterprise journey and with unstructured documents. The pain of pulling data out of PDFs, invoices, scans is big. Even if you use the best of the AIS, be chat, GPT or cloud, you’ll find certain errors in that. Why template based OCR breaks when layouts changes? We all know it because see on one of the structures it

00:31 performs very well but on the other it just fails completely. So today I would be presenting you one of the best tools for unstructured data extraction that is unstruct as the name suggests it’s used by multiple enterprises and yes it’s great to use it use lms to extract structured JSON from document like PDF images and scan. You define what you want to extract in natural language and boom you get your results within seconds and use any LLM

01:03 use any vector DB of your wish. The entire platform is built in such a way that you can plug and play multiple things. The opensource edition is AGPL and is available as an API ETL pipeline as well. So I’ll be starting off to show you how you can use it. I have signed into unstructed cloud and this is the platform that looks like this is a prompt studio that we have got right now and this is another document that is present here as you can see this kind of

01:34 uh bill that you can see here so this is the first question that I’ve asked what is the name of the issuer or the bank who has issued this credit card and here you can see that token consumed was 4,700 cost was this and extraction was blue cash from enter American Express. What is the name of the customer to whom this credit card state belongs to? Now let’s hit enter and you’ll see within seconds you would be able to get your solution. Joseph

02:05 Pollson. So it is able to give you exact replies as you want to have. So there might be multiple addresses in the context provided for you. First gather all addresses. For the customer address form a simple JSON object with the following fields. full address, city, zip. Let me hit enter for this also. This will take a while. And yes, there’s a structured output that we have got. Now, time consumed is just 3 seconds. And this is not just

02:35 that you think that we are just using G40. You can go here. You can change out multiple things. You can change the LLM as well. You can add a new LM profile of your choice. You have summarizer vector also. So you can select an LM here too. LM challenge custom data grammar preamble post amble highlighting every single feature is present here and it’s crazy work done here. I can test you out with multiple other documents too. Handwritten form let’s test it out. I

03:07 think for uh printable documents for printed PDFs it’s quite easy but here now you can need to check out how it works with handwritten documents too. Here you can see that the first name is a little difficult. I am a card holder. Some dates are present. It’s half uh filled by hand. Half is printed. Let’s see what it generates for us. What is the name of the applicant? So not just on printed PDFs and all. It performs well on handwritten documents also and

03:40 which most of the time can be very helpful for you for structuring out your unstructured data. This is might taking a small while. Just wait for a while now and you should be able to get an answer for this particular PDF as well. Cover 021 document. I am a card holder. As you can see, it is able to extract that out. Now you can also extract out multiple things. What is citizenship of the applicant? Respond is one of the following. US citizen, PR alien, nonpr

04:12 alien, US citizen. So even whatever is the text written it is able to answer in your format too. Form a JSON object with the following information. We’ll hit enter here again. Uh it has started working now. It it is also giving an entire analysis the cost time token etc. You’ve got the desired output here too. So quite a great tool for everyone to use. uh and not just that talking about the last thing as I

04:45 showed it was model agnostic also you can change multiple things you can deploy it as an API too I click on deploy as API proceed sample credit card passing project API deployment name is this let’s hit enter minimum maximum 30 characters allowed let me name it as test next tool settings you You can enable everything here. Let me enable this. I’m using GP4 LM free trial and create

05:17 the deployment. Now you can use it as an API or ETL pipeline too. View API deployment and yeah here you can see there is a workflow API endpoint is this and you can start using it ASAP. Now if you look at the code snippet how to use it just need to click here. You can mention your information. There’s the API key and you can get started with using the parsing here too. So this is a great platform for enterprise uh unstructured document

05:50 generation and my friends this is a banger of a platform that you should definitely try out if you are into unstructured data arena. With this it’s a wrap. I hope you try it out. The link is in the comment section below and do click it out work on it. Let me know in the comment section how you feel about the tool. Thank you so much.