Ready-to-Use APIs for Instant, Domain-Specific Document Automation

[00:00:00] 

Hey, everybody. Good. Uh, hey everybody. Hope all of you’re doing well. I’m Mahashree from Unstract and I will be your host for today. So as you can see today, we’ll be talking about ready to use APIs for instant domain specific document automation. Now if you’re wondering what this exactly is, let me rewind a little and give you a bit of context.

So over the last couple of months, large language models have taken over the document 

[00:00:30] 

extraction space. They are bringing transformations that are completely changing the way things are being done. However, one of the biggest observed strengths of elements in document extraction would be the ability to shorten document extraction development cycles.

So this is one of the biggest strengths, and now imagine if these cycles could be cut down even further. So that is what we are looking to achieve with ready to use APIs. So essentially in this session we’ll be taking 

[00:01:00] 

a look at Unstract’s new offering, the API Hub, where we have a collection of ready to use APIs for different document types.

So what we see ideally is there are businesses that process the same documents more or less. So you’d see all the businesses almost process invoices, purchase orders, bank statements. So how can the development cycles for these documents be brought down? So that will be the core discussion of today’s session.

We’ll start off the session with what is 

[00:01:30] 

API hub to give you a formal introduction. And then we’ll move on to covering a comparison between API hub versus building from scratch. So how do both of these approaches work to document extraction? What are their benefits or disadvantages? And in what use cases?

Would either of these approaches be a better fit? Moving on from there, we’ll take a look at a live demonstration of the API hub in Untract, and we’ll, uh, see how this can be deployed in real time. And finally, 

[00:02:00] 

in case there are any questions remaining, we’ll also be venturing into a live q and a session where you can ask questions to, uh, one of our experts who will be moderating the question and answer segment.

So that said. Here are a few session essentials that I’d like to quickly run over before I get started with what is API hub. So firstly, all attendees will be in listen only mode. I mean, this webinar is in listen only mode, so all, all attendees will be on mute. Secondly, you can post your questions in the q and a tab at any time 

[00:02:30] 

during the session, and our team will be able to get back to you with the answers via text.

Have any, uh, technical glitches that you face during this session. You can always, again, post them in the chat tab and you can use the chat tab to interact with fellow attendees as well. And finally, when you exit this webinar, you’ll be redirected to the, uh, feedback form that we have attached. And, uh, we’d love it if you could leave your feedback as it would really help us improve our webinar experiences going forward.

So that said, let’s 

[00:03:00] 

kickstart with the first discussion. That is, what is API hub? Now API hub is essentially a collection of ready to use APIs in abstract that are carefully designed to extract industry specific or domain specific data from common business documents. So we’re talking about documents like invoices, purchase orders, bank statements, and uh, credit card statements.

So these are all documents that almost all businesses incur. How can the development cycles for document 

[00:03:30] 

extraction be, uh, made shorter because. Ultimately we are, uh, most of these businesses, as we see, even with our customers, are looking to extract the same data from across all these different documents.

So this, with the API hub, we see that we can skip lengthy development cycles, and this is essentially a plug and play service where you do not really have to, uh, do any of the heavy lifting. It’s already done and made ready for you. So essentially what you have to do is take this API and plug it into your workflow 

[00:04:00] 

wherever relevant and you’re good to go.

So what are the differences between the API hub versus building from scratch, which would uh, be the traditional approach? So firstly, we have speed, which is one of the obvious benefits that we can see because when the development cycles shorten the time to go until you get your output data also shortens, and this again contributes to increased speed of document extraction.

Secondly, we have lower development effort. So as I mentioned earlier, the heavy 

[00:04:30] 

lifting is taken care of by the ready to use API. So this means that your dependency on the IT team is brought down to a very large extent, and this also makes these APIs much accessible to your teams. Thirdly, we have reduced maintenance.

So with plug and play APIs, Unstract takes care of the updates, the improvements, uh, that are made to the extraction logic. So this helps you avoid any technical debt since APIs are tested across a wide range of use cases and conditions before they are 

[00:05:00] 

deployed and made available to you. So this, again, brings down the maintenance to a very large extent.

Moving on, there is domain specific extraction. So our team has done the research work for you and we’ve built the APIs to capture all the critical data points that are usually looked to, uh, extract from each of the document types. So depending on your domain, depending on your industry, we have a list of all the key data.

So you do not miss any key fields, even if you were to develop them on your own. 

[00:05:30] 

Going on from moving on from there, we have compliance and accuracy. So our APIs follow compliance standards and are designed, uh, to minimize errors and give you greater confidence in the results that you get. And finally, there is the one-stop solution.

So now imagine if all these benefits are given across all common business documents in one place. So that is what your API hub entails. So does this mean that APIs can completely replace building from scratch, or can 

[00:06:00] 

APIs completely replace the development that we have in terms of using LLMs and entering prompts to extract data?

Well, that is also not the case. What we are seeing right now is two different approaches to document extraction and this, these two approaches are equally beneficial to different use cases. For instance, as we saw, as we talked about earlier, if you are dealing with common business documents, then the API hub would make more sense because it’s already, it already has done the work for you.

However, if 

[00:06:30] 

your business requires, uh, you to extract data from. Uh, documents that are very specific to the business or, uh, you have, you want to customize the projects and have the control in your hands, that is when you’d probably go for the traditional development. So with traditional development, you have much tighter control over your projects.

You can always go back and edit the extraction logic as you want. Whereas with the API hub, if you are using an API from there, you’ll have to wait for the team to probably get back and make the, uh, 

[00:07:00] 

changes for you. So in terms of customization and having tighter control over your projects, uh, if you want to extract documents that are very, very specific to your business, or you want the extraction format to be in a, in such a way that it.

Probably only works for your business, then that is probably where the traditional development, uh, gives you much greater advantage and is probably the better fit out of the two approaches. So that said, let’s get to the core of what this is. We’ll move on to the interesting chunk of this webinar 

[00:07:30] 

where we’ll see the API hub in action.

However, I do notice that some of the attendees today are completely new to untraced. So let me take, take a quick detour and I’ll take you through what is untraced, what we do, and uh, basically all of that so you get a bit of context before we move on into the demo. So Unstract is an LLM powered unstructured data, ETL platform.

If I had to, uh, give you an overview of, uh, the various capabilities in the platform, I’d probably segment it 

[00:08:00] 

into three categories. So up first we have text extraction, which is one of its key capabilities. So Unstract extracts, layout preserved text from all your uploaded documents. This is done using a text extractor tool.

You can connect with various text extractors, text extractor tools within the platform itself. However, one of our, uh, sought after tools or text extractors would be LLM Whisperer, which is again, Unstract’s in-house text extractor, and these two go pretty well hand in 

[00:08:30] 

hand. And LLMWhisperer is known to generate LLM ready outputs.

It’s. The secret sauce behind the tool is basically, it, uh, preserves the layout of your original text, and, uh, it also converts it into a format that is LLM ready. And this is also a tool that is available as a standalone solution, depending on our customer’s needs going on from there, once you extract the raw text from your documents, you’ll have to send that context.

L And what does the LLM do on this context? That is 

[00:09:00] 

where the development phase comes into the picture. So in the development phase, you will be creating prompts in a prompt engineering environment called Prompt Studio, where you specify, one, what data you’re looking to extract, and two, what is a schema of extraction that you’re looking for.

So these two are the key points that you should look to specify across all the extraction items when you write your prompts. And you can test these prompts across multiple document variants as well within a given prompt studio project. And you also have access to various 

[00:09:30] 

accuracy enabling capabilities, as well as cost saving capabilities in the prompt studio.

We will, uh, be going over this a little bit in this webinar, and once you have. Created a project in this from studio and you’ve tested across different documents, you can then deploy this project in any of the deployment options that we have now, what we are trying to achieve with API hub is that.

Spending time in the Prompt Studio project is actually one of the core chunks of document extraction, uh, when you’re trying to do it, using 

[00:10:00] 

elements to actually enter the prompts and test how it works across different documents. So this entire core chunk will be replaced by a ready to use API because you do not have to do this entire work that is required in the development phase with the APIs we have in the hub.

So that is what we are looking to achieve, and once you have your APIs ready, you can deploy it either as an API deployment as an ETL pipeline, task pipeline, human in the loop deployment, and for certain advanced deployment requirements. We also support a 

[00:10:30] 

for agent pick workflow automation tools, and, uh, we also support an abstract MCP server as well as an MMCP server.

So that said, here are again a few numbers that would throw light on the platform. So currently Unstract has over 5.5 K stars on GitHub, a 950 plus member Slack community, and around 8 million plus pages that can, that are currently being processed by paid users alone per month. So that said, um.

Here are the different, 

[00:11:00] 

um, additions that Untract and LLMWhisperer come in. So, Unstract is also available as an open source offering with limited features for you to, um, try it out on your own and experiment with it. And we also offer a cloud offering as well as on-prem version. For LLMWhisperer again, there is an LLMWhisperer playground where we offer a generous, um, a hundred pages that you can, uh, process using LLMWhisperer and text, uh, extract text from for free for the first a hundred pages per day so you can, uh, 

[00:11:30] 

deploy end-to-end capabilities on these pages and see how it is working in the context of your own documents.

And we only, we will only be charging you go beyond that limit. LLMWhisperer again, can be deployed on the cloud or as an API, as a Python client, JavaScript client, and again, as we looked earlier, as an innate node or an MCP server. Again, both these platforms are, uh, compliant to various compliance standards like ISO, GDPR, SOC 2, and HIPAA.

So that said, let’s move 

[00:12:00]

into the demo and we’ll, I’ll quickly take you around the platform and what we have in store. So as you can see, this is the unstuck interface. This is how it would look once I open. Once you open the application, what, where you can find API hub, and let me get to that first is probably over here on the top left, uh, handset toggle.

So we have unstuck, and then you can also access LL and Bisra. And now newly we’ve added the, the third service that is the API hub. So when I click on the 

[00:12:30] 

API hub, you can see that we have all the APIs currently available over here. So as of now we have six APIs. These are a mix of utility APIs as well as APIs that are designed to directly extract data from documents.

And we will be adding more APIs to this space. So it’s, uh, good to keep a good watch on this space and, uh, we’ll anyways be updating you with the addition of new APIs as well. So, uh, utility APIs. Over here, uh, would be the all table extractor and the PDF splitter, API. So 

[00:13:00] 

these two APIs, what we mean by utility APIs, for those of you that may be new, is that they’re designed not to directly handle the extraction of data from documents, but these are APIs that handle the supporting services that will aid in the extraction.

For instance, if I take a look at the PDF splitter. This is a utility, API that takes in bulk or multi-page files and breaks them down into individual well-structured documents. So you might be handling documents in your businesses that contain 

[00:13:30] 

various, uh uh, I mean various other documents in one large file.

So in order to process these documents separately, you might be currently going about this manually, or you might have certain, uh, uh, chunking or indexing rules that you have set up. However, with the PDF splitter, it automatically identifies what are the different documents within a larger file. And it is able to segregate them and give this, give them to you in the order in which they appear in the original file.

So for example, let’s say you have around 200 page loan 

[00:14:00] 

application with KYC forms, financial statements and agreements. Then the document splitter is able to automatically split all these individual files and give them to you separately. So this is what we are looking to achieve and this API can be especially useful in industries like finance, banking, legal, or healthcare.

So, as you can see in the API, we have an overview over here that gives you a better understanding of what this API does, what are, what are its key features, and uh, what are the different use cases in 

[00:14:30] 

which this can, you know, come to use. So you can go through this to learn more about the PDF splitter. A API.

And you can see that we also have the data extracted tab where we get an idea of what is the data that will be extracted from this API. So we are extracting some, uh, metadata for each document, like the, uh, the document type classification, confidence, coding, and so on. And we also have the individual PDFs, uh, that are given to you with the quality of the original document preserved.

We have intelligent naming of the PDFs and they’re 

[00:15:00] 

also organized in a logical ordering, uh, order according to, uh, how they appear in the original document. And we also support batch processing to handle multiple documents, uh, at the same time. And again, there is the API response tab right next to the data extracted tab.

Over here you can see a, a sample preview of how your extracted data would look. So what are the different, um, what is the JSON structure that, uh, this API would give. So you can go through this in detail as well. 

[00:15:30] 

So that said, uh, let me. Upload a sample document and we can test this API. So you can see that you can actually try this API in the playground or, uh, which, which is a cloud offering that we have.

Or you can also download the Postman Collection and deploy it locally on your Postman account. So let me just try this in the playground, I’m going to be uploading a sample document, so we’ll upload a document. Uh, which has a number of, uh, other, I mean it, this document basically has a 

[00:16:00] 

lot of individual documents that are combined together.

So while this extracts the data, let me show you what this document really is and we’ll come back to the extraction.

Alright, so this is the document that we are dealing with. This, uh, document, as you can see, contains a credit card statement To start off with. And then we have a loan application that is also achieved. I mean, there’s also, uh, attached and uh, you can see that we have different sections within the loan application, 

[00:16:30] 

like the borrower information, um, assets and liabilities and so on.

So this will go on for another three, four pages. And we also have some, uh, terms and conditions declarations. So once all of, uh, I mean this, all of this contains, uh, is contained within the, uh, loan application and. Thirdly, we have a purchase order, I mean an invoice that is attached to the document as well.

And finally, we have a purchase order. So this, uh, could, I mean, it’s a very real 

[00:17:00] 

example. You could have different documents put together for logical reasoning, or sometimes they just end up together and you have to segregate them. So, uh, this is what we’ve uploaded so far. Let’s see how the extraction is done on the platform.

So you can see that I got a message. The document is split successfully, and I, I can download these split documents over here. So let me just open this and, uh, it is downloaded as a zipped file and you can see the different documents individually separated. So right from the credit card statement, if I 

[00:17:30] 

open this, this is just the credit card statement that you had seen in the earlier document.

And, uh, the financial, uh, statement is right here. And we have the loan application over here and similarly. We also have the invoice as well as purchase order. So this is what you can achieve with the PDF splitter, and you also have a JSON with all of the details of the extracted documents combined.

So you have details like the file 

[00:18:00] 

name, the start page, the end page number, uh, the header, as well as key entities that were found within the document for across all the extra, uh, all these, uh, segregated documents that we have downloaded. So this is again, of uh, great use to industries like, uh, finance and healthcare as I’d mentioned earlier.

So with that, we’d seen the first utility, API. Now let me, uh, go and explore the other utility, API that we had. That is the all table extractor. 

[00:18:30] 

Now the all table extractor is an API that is used to detect and extract just the tables from documents. So, uh, certain documents like invoices in finance or accounting expense reports, or if you’re looking at supply chain, then you might be dealing with packing lists, custom declarations.

In insurance, this could be, uh, claims forms or in healthcare, it could be hospital bills, medical test reports. So all of these documents many times contain a lot of tables 

[00:19:00] 

that, uh, have the core information that you’re looking for. And sometimes you just want to extract just the tables from the document and use the data from the tables for downstream operations.

So in this case, you can then deploy the all table extractor, which automatically identifies tables within a document and extracts them for you. So again, again, going through the overview, we have a bunch of key features like, um, universal table detection, structure preservation. So it maintains the relationship between the headers and the cells.

It has multi. Multi-format 

[00:19:30] 

output. So you can either, it returns data as j uh, JSON or TSV. And we also support batch processing, like, uh, to handle multiple documents efficiently. And you have a bunch of use cases that this might come in use, uh, for and supported input formats as well as stable detection capabilities that you can go through.

And in the data extracted, you will basically have all the tables that you have extracted. And, uh, this is just a preview of how the output would look. So let me try this again. I would be uploading a sample document 

[00:20:00] 

with a couple of tables and let me try extracting the data from this document. So while this extracts again, let me take you through the document.

Uh, this is a climate and environment analysis report for 2024. So you can see that we have around five tables in this document with different information on the climate, like the global temperature trends, uh, weather events, energy adoption, climate policies, and the environmental outlook for 2025. So let’s see how this 

[00:20:30] 

data is basically being extracted by the, uh, API.

So ultimately this is the kind of output we’re looking to receive, uh, with a JSON view as well as the TSV view, which can be downloaded individually. So you can see the various data that has been, uh, extracted over here. So this output can be then deployed into downstream, uh, operations. Let’s move on to covering the bill of lading, which is one of the, uh, APIs that we have 

[00:21:00] 

to directly extract, uh, doc document data from the documents itself.

So the bill of lading is basically a legal document in shipping and logistics to prove that goods were received or transferred. So here we have an overview of what, uh, the bill of leading is and what this API basically does. We have a look into the, uh, data extracted by this API. So a typical bill of lading would have details like, uh, the document number, the issue 

[00:21:30]

date, uh, BOL type, bill of lading type, and so on.

And it would also have the details of the parties involved, like the shipper, consignee, and uh, carrier. The transport details as well as the routing details, the cargo details, financials charges, and so on. So we have all, we have quite an extensive list of data fields that we’re looking to extract from any given bill of lading.

And, uh, in the API response over here we have, um, a structured JSON with all the different output data 

[00:22:00] 

fields extracted. So we have the names of these data fields along with the output data type in which they’ll come. So this is how your response would ideally look. So let me, um. Upload, upload a sample bill of lading and we’ll extract some data from this document and we’ll see how this works.

So meanwhile, let me also show you, uh, what this, uh, document looks like. So as you can see, we have the shipper details, the consignee details, the carrier vessel details. 

[00:22:30] 

The port of loading and discharge and a couple of other conditions. So this is a document that I’ve uploaded right now and let’s see how the text extraction results, um, appear.

Alright, so we have the data extracted over here. So you can see the structured JSON with all the, uh, output fields. So what, whichever output field was present in this particular document, you can see them extracted, the others are returned as null. So, uh, you can go through an extensive list of data items 

[00:23:00] 

that can be deployed in downstream operations.

So this is available as a JSON view, as well as a TSV view. And, uh, you can download both these, uh, different, uh, output types as well. So this said, uh, I think we’ll also go through one more, uh, ready to use API before, uh, we conclude the session. So let me take the bank statement extraction API, so you can see that we again, have the overview in this, uh, bank statement, uh, extraction.

API We are looking to extract the account 

[00:23:30] 

holder information, the account number, and uh, bank details, transaction history with the dates, amounts, and descriptions. The opening and closing balances, as well as the statement period information. And, uh, you also have an idea of how the, uh, API response would look.

So since we have already tried the playground for the API hub, I thought we can, you know, download the Postman collection in this case. Now I’ve already, uh, set up the, downloaded the collection, and I’ve set this up on my Postman account. So, uh, what you see over here is the 

[00:24:00] 

bank statement extraction API.

So over here we have a postman request that we have sent where we’ve uploaded a bank statement. So, uh, let me just send this again to run this.

[00:24:30] 

So let’s just wait until this has been, um, co completed. So you can see that the task has been queued for, uh, text extraction. So the document that we are dealing with for the bank statement over here is, uh, this is the document that I’ve uploaded. So we have some bank details, the account holder information, the transaction history that we have, and a couple of other details.

So let’s just wait a couple of seconds until the do, uh, data has been extracted. 

[00:25:00] 

So let, let’s explore some of, uh, the other two APIs that we have over here. So we have the invoice extraction, API. So invoices are one of the most commonly ex, uh, processed documents. So we, uh, you, I mean, you can always. Come here and check this API out.

And it’s very similar. Upload your own document or you can see the sample results with some of our pre uploaded documents over here. And again, you have the option of downloading it as a postman collection. And similarly, we also have another API for purchase orders, uh, extraction.

[00:25:30] 

So you get, again, an overview, the data extracted given over here.

This is again, another extensive list that you can go through. And you again have the API response. Uh, let’s just go back and check. Postman, if this has been completed. So I have a status, API as well that I can, uh, uh, call and you would get a status on what is the status of the extraction. So finally, when I retrieve the values, you have the values over here.

So we have the account, uh, holder information, the number, the 

[00:26:00] 

account type. So we have savings and, uh, various other, uh, data that is usually used or extracted from bank statements that we have. So this is basically how you can deploy the API hub and, uh, use it, use this data in downstream applications. So with that, we’ve seen how the API Hub works and, uh, let’s just quickly compare as to what this really means when we are comparing it with traditional development methods.

So, going back to the platform, 

[00:26:30] 

uh, if I have to show you what this means, so we have the, let me just move into Untraced. And they’ll show you, um, how the traditional development will look for the different, uh, document extraction use cases. So this is how the Untract interface looks, folks. And if you are signing up for the first time, what you’ll have to do is set up certain prerequisite connectors for the platform, uh, to run and for you to smoothly, uh, develop your projects.

So again, we have a bunch of lms, 

[00:27:00] 

popular LLMs in the market that you can connect with vector dbs, embedding models, as well as text extractors. So here’s where you’d also find LLMWhisperer as a text extractor that you can choose, uh, from the various other options that we have available. So once this is done, you are ready to start developing your extraction projects, which is, as I mentioned earlier, done in Prompt Studio.

So in this webinar, for the want of time, I would be getting into an existing project that we have, which is a credit card parcel. So you can 

[00:27:30] 

see that we have a sample credit card statement over here. Along with a couple of prompts that we have given to extract specific data from, uh, credit card statements.

So you can see that each of these prompts are quite detailed and, uh, it gives you an, uh, proper, I mean, it’s important to structure your prompts in such a way so that you are extracting exactly the data that you want and also in the format that you want. So you can specify formatting details, like format, the name of.

Uh, we are extracting the customer name over here, 

[00:28:00] 

for instance, so you can, uh, specify details like formatting the name of the customer with the first letter of each name to be capitalized. And you can also give instructions on how the output JSON should be, uh, formatted if you are looking to extract your, uh, data in the form of, uh, JSON.

So we have various other output data types. That you can choose from. So this is the process that we are essentially skipping when we go for APIs from the hub. So, um, you can see how this has detailed, uh, 

[00:28:30] 

descriptions of what is the data that we are looking to extract. And in this case for, um, since this is a sample project that I’m showing you, we just have for, uh, data items that we are extracting.

But for any. Real-time business use case, you might have close to 10 to 30 data fields that you want to extract. So for each of these data fields, you’d again have to go into the prompt engineering and specify the data and the format and all these different details. So this effort can be spared, especially when we are dealing with.

Common business 

[00:29:00] 

documents because as I mentioned earlier, businesses are looking to extract the similar data, uh, data fields from, uh, common documents like invoices or bank statements, purchase orders, and so on. And again, in the prompt studio, you can test these prompts against multiple document variants.

So you can see over here within. Project I have uploaded, uh, three credit card statements overall, so if I have to shift to another credit card statement, you can see how we have a different statement over here in the preview, and we 

[00:29:30] 

have a different customer name extracted as well. So this is how you can, again.

Test, uh, how your documents, I mean, how your prompts are working against different document variants. And this effort as well has been taken care of with the API hub, as our team has already rigorously tested the APIs across different variants for each of the document types. So this is, uh, basically the, uh, this is how visibly we can see the, uh, difference or the benefit of using APIs.

Plug and play APIs for common 

[00:30:00] 

business document types. So again, within the, uh, prompt studio, we do have advanced settings that will help you closely control, uh, how your data is extracted. You can, uh, you have accuracy enabling capabilities like LLM challenge and, uh, certain cost saving features like summarized extraction as well.

So I wouldn’t be getting into the details of these features in this webinar. We do have an extensive documentation that you can take a look at for those features and, um. We, we also have previous webinars 

[00:30:30] 

that you can take a look at as well. So with that, folks, we saw how the difference between deploying, uh, APIs from the Prompt Studio and using the readymade plug and play APIs that we had in the API hub.

So the final step in Prompt Studio would be to click on deploy as API. And finally, this is the deployment that you can then take and plug it into your workflows wherever relevant. So this is the effort again, that we are saving. APIs. So that said, let 

[00:31:00] 

me uh, move back to my presentation and. Um, so we saw in this webinar how you can, uh, use the various, uh, plug and play APIs for common business document types and, uh, prompt studio projects would be more suitable for the, uh, use cases where you might want to customize the documents and, uh, customize the extraction from your documents and, uh, you know, have tighter control over.

How the extraction logic is altered. So if you would like to set up these 

[00:31:30] 

deployments for your business, you can always book a free demo with one of our experts where we will be able to sit with you one-on-one, understand your needs, and see, uh, which kind of, um, development approach would probably suit your particular business use case.

So you can register for it using the links that you would find in the chat. And in case we have any questions, we can also venture into a q and a.

[00:32:00] 

It looks like we do not have any questions. So thank you everybody for joining this session today. Uh, we hope to see you at our upcoming evenings as well. Thank you.