Imagine you need to find a specific clause buried in a 200 page legal contract. You don’t read every word, you skim the table of contents, jump to the right section, and scan for what you need. PageIndex teaches AI to do exactly that. And in doing so, it’s quietly dismantling one of the most deeply rooted assumptions in AI document retrieval.

First, What Is RAG?

Large language models like GPT or Claude are incredibly capable, but they have a fundamental constraint: the context window. They can only process a limited amount of text at once. Feeding an entire 300 page annual report into a single prompt is often impossible, and even when it works, model performance degrades as the document grows longer. Researchers at Chroma have documented this as context rot a phenomenon where accuracy quietly collapses as context length increases.

Retrieval Augmented Generation (RAG) was the standard fix. Instead of feeding the whole document, you extract and inject only the most relevant pieces before the model generates its answer. Think of it as giving the AI a tailored cheat sheet for each question.

The classic approach works like this:

• The document is split into fixed size text blocks say, every 512 tokens.

• Each block is converted into a numerical vector using an embedding model.

• All vectors are stored in a vector database.

• When a question arrives, the question is also embedded, and the database finds the most similar vectors.

• The top matching chunks are fed to the LLM alongside the question to produce an answer.

Simple, elegant, and for many use cases effective. But this model has serious cracks when applied to complex professional documents.

Where Traditional Vector RAG Breaks Down

The core assumption of vector RAG is that semantic similarity equals relevance. Find text that sounds like the question, and you’ll find text that answers it. In casual or general purpose retrieval, that’s often true. In dense professional documents, legal contracts, pharmaceutical trials it falls apart fast.

Query intent doesn’t match document content

When someone asks “What are the deferred tax assets?”, that question doesn’t share high vector similarity with the specific table buried in Appendix G that actually answers it. Queries express intent; documents express content. They live in different spaces.

Everything sounds alike

In a financial filing, dozens of sections will use near identical language about revenue, risk, and liabilities. Vector search returns what sounds right not what is right. The top-k results can be confidently wrong.

Hard chunking destroys meaning

Slicing a document every 512 tokens is arbitrary. Sentences get cut in half. Tables get split across chunks. Footnotes get separated from the figures they explain. This silently corrupts retrieval quality in ways that are hard to diagnose.

No memory between questions

Each query is treated in isolation. If you asked about “revenue” and then ask “how does that compare to liabilities?”, the retriever has no idea the follow up is related. Every question starts from scratch.

Cannot follow cross references

When a document says “see Table 5.3” or “refer to Appendix G”, vector search has no mechanism to follow that cue. Those phrases share no semantic similarity with the data they point to. The reference is invisible to the system.

The result? Even highly optimised vector RAG pipelines typically achieve low accuracy on complex document Q&A tasks. Not because the technology is bad but because it was designed for a different kind of problem.

Enter PageIndex

Developed by Mingtian Zhang and Yu Tang at VectifyAI (September 2025), PageIndex throws out the vector database entirely. Instead of asking “which chunks sound most like this question?”, it asks something far more human: “if I were a domain expert, which section of this document would I go to first?”

The system was inspired, in part, by the tree search strategies behind AlphaGo the idea that intelligent navigation through a structured space can massively outperform brute force pattern matching. Applied to documents, this means building a navigable map of the content, then reasoning over it.

Phase 1: Building the Index

Before any questions are asked, PageIndex reads the document and builds a hierarchical JSON tree structure essentially a machine readable Table of Contents. Each node in the tree represents a logical section: a chapter, a sub section, a page range. Each node gets a title, page numbers, and a concise summary of what it contains.

The result is a compact index that lives inside the model’s context window not in an external database. The team calls this an in context index: a structure the model can read, reason about, and navigate in real time.

Phase 2: Reasoning Through It

When a question arrives, no similarity search happens. Instead, the LLM reads the index and reasons about where the answer is likely to live then goes and gets it. The loop works like this:

1. Read the tree. The model reviews the full hierarchical index to understand the document’s structure.

2. Pick a section. Based on reasoning not similarity scores it selects the node most likely to contain the answer.

3. Extract what’s there. It reads that section’s raw content and pulls out anything relevant.

4. Enough? Answer. Not enough? Loop. If more context is needed, it returns to the index, picks another node, and repeats including following cross references like “see Appendix G”.

5. Generate the answer. Once sufficient information is gathered, the model produces a grounded, traceable response.

A real example

Imagine you ask: “What is the total annual revenue?”

The report has 90 pages. On page 12, there’s a section on Q4 performance that mentions revenue numbers. So vector RAG grabs it, since it sounds relevant. But that page only covers the last quarter, not the full year. Somewhere in the middle of that page sits a quiet little line: “For full year figures, see the Annual Summary on page 67.”

Vector RAG ignores it. That sentence shares no semantic similarity with “total annual revenue,” so it’s invisible to the system. The AI returns the Q4 number with full confidence. Wrong answer.

PageIndex reads the table of contents first, spots a section called “Annual Summary” on page 67, and reasons that a full year revenue figure is exactly the kind of thing you’d find there. It navigates directly to page 67 and returns the correct number.

The Honest Limitations

PageIndex is genuinely impressive for its use case. But it’s not a universal replacement for vector RAG and the team is upfront about the trade offs.

Latency and cost

Every retrieval involves multiple LLM calls: read the index, select a node, extract content, decide if it’s enough, loop. Traditional RAG answers in milliseconds. PageIndex can take several seconds per query. For consumer scale applications where volume is high and response time matters, this is a significant constraint.

It doesn’t scale to large document collections

PageIndex thrives on deep analysis within a single document. When you need to search across thousands of documents simultaneously, building a tree index for each and reasoning over all of them becomes impractical. The system falls back to vector search in these scenarios and loses its edge.

Still maturing in production environments

While the architecture is elegant, the system is still in the “assess” phase for enterprise adoption worth experimenting with, but not yet battle tested enough to carry critical production workloads at scale.

It trades one approximation for another

Vector RAG relies on embedding approximation. PageIndex relies on LLM reasoning approximation. The model can still navigate to the wrong section, misread a summary, or overlook a sub node. The failure mode shifts it doesn’t disappear.

Requires well structured documents

The tree index is built from document structure. If a document lacks clear headings and logical organisation a poorly formatted scanned PDF, for example the quality of the ToC degrades, and so does retrieval accuracy.

Benchmark scope

The flagship 98.7% result comes specifically from FinanceBench a high quality financial Q&A dataset. Performance on general purpose, multi domain, or highly ambiguous question types across diverse document formats remains less thoroughly evaluated in independent studies.

Which Should You Use?

Use PageIndex when…

You need deep, precise extraction from a structured document, financial filings, legal contracts, medical records, technical specifications. Accuracy is non negotiable. You can tolerate higher latency and cost. Try it at chat.pageindex.ai or explore the open-source repo.

Stick with vector RAG when…

You need fast, scalable semantic search across a large collection of documents knowledge bases, customer support, content discovery. Sub second responses are required, and cost-per-query needs to stay low.

The emerging best practice noted by multiple researchers is a hybrid approach: use vector retrieval to narrow 10,000 documents down to the top 5 candidates, then deploy PageIndex for precise extraction within those finalists. Best of both worlds.

The Bigger Picture

PageIndex is part of a broader shift happening across AI research, from static, embedding based retrieval toward agentic, reasoning based systems. The same instinct shows up in Graph RAG, Self-RAG, and a growing body of work arguing that LLMs are capable of navigating knowledge structures, not just pattern matching against them.

Even Anthropic’s Claude Code has reportedly moved away from vector databases for code retrieval in favour of reasoning based approaches (discussed here) — suggesting this isn’t a niche experiment, but a signal of where the field is heading.

The vector database market is currently valued at $2.2 billion and projected to reach $10.6 billion by 2032. PageIndex doesn’t kill it , it contextualises it. For precision critical, document heavy domains, the message is clear: thinking beats searching.

Imagine a world in which machines don’t just speak our language, but can use it with the brilliance of an artist or the charm of a master storyteller. This isn’t science fiction; it is the reality, as we step into with Large Language Models (LLMs). Leave behind hard-coded and predetermined functions — these AI wonders learn and adapt, transforming the boundaries between manmade creativity.

Large Language Models (LLMs) represent a subset of deep learning and stand at the intersection of generative AI, showcasing the ability to produce diverse forms of content, including text, images, audio, and synthetic data. These models are characterized by their dual-stage training process, where they are initially pre-trained for general language understanding and then fine-tuned to cater to specific tasks. Their versatility allows them to address a spectrum of common language problems across industries, such as text classification, question answering, and document summarization. Unlike traditional machine learning models, LLMs leverage their immense size, both in terms of training datasets often reaching the petabyte scale and the high count of parameters or hyperparameters. This unique combination empowers LLMs to offer efficient solutions for various applications, making them adaptable to different domains through fine-tuning with relatively smaller datasets.

LLMs have several distinctive features that make them stand out in the field of natural language processing. Most importantly, their scale is impressive: trained on data that can reach petabyte order and with a high parameter count, they reflect how comprehensive the model’s understanding of reality has become. In addition, LLMs are highly general-purpose tools capable of handling a variety of commonly encountered language tasks because natural human language is universal across domains. Nevertheless, because of resource limitations, their training is limited to specific organizations. These models are already pre-trained on large data for general purposes and then fine-tuned with small but specific datasets to adapt the model appropriately. When it comes to advantages, LLMs are unparalleled when used for different tasks such as translation transcription and classification among others along with being efficient despite using limited field training data because they enjoy decent performance in terms of few-shot and zero-shot cases. Furthermore, LLMs show progressive improvement because their performance increases as data and parameters keep growing; for example, Google’s PaLM is a giant model with 540 billion parameters. LLMs represent a drastic shift, from rigorously specified programming paradigms to more generative models that learn by example and may eventually create their content.

The differences between LLM development and traditional machine learning can be accentuated by comparing the two processes, as less amount of expertise is necessary for developing an LLM beyond prompt design compared to considerable resources such as training examples used in ML which also require time along with hardware devices.
As someone fascinated by the potential of LLMs, I’m excited to see how this technology shapes our world. What are your thoughts?

Clean code is a term used to describe a codebase that is easy to read, understand, and maintain. Writing clean code is essential for any software development project because it ensures that the codebase remains manageable and efficient over time. In this blog post, we will explore what clean code is, why it is important, and some best practices for writing clean code.

What is Clean Code? Clean code is code that is well-organized, easy to understand, and straightforward to maintain. It follows a set of best practices that ensure readability and maintainability. Clean code should be self-explanatory and should communicate its intent clearly to the reader.

Why is Clean Code Important? Clean code is essential for a variety of reasons. It makes it easier for developers to understand and modify the codebase, which leads to faster development and fewer errors. Clean code is also easier to test and debug, which reduces the amount of time required to identify and fix bugs. Additionally, clean code is more maintainable, which means it will continue to work effectively as the project evolves over time.

Best Practices for Writing Clean Code

1. Use Meaningful Names: Use descriptive and meaningful names for variables, functions, and classes. Avoid using vague or ambiguous names, and choose names that accurately describe the purpose of the code.
2. Keep Functions Short: Keep functions short and focused. Functions should ideally do one thing and do it well. If a function is too long or complex, break it up into smaller, more manageable functions.
3. Write Code that is Easy to Understand: Write code that is easy to understand, even for someone who is not familiar with the project. Avoid using complex or convoluted logic and try to use straightforward, easy-to-follow code.
4. Be Consistent: Use a consistent coding style throughout the project. Consistency helps to make the codebase easier to understand and maintain.
5. Use Comments Sparingly: Use comments sparingly and only when necessary. Comments should be used to explain why something was done a certain way, not to explain what the code does.
6. Keep Code DRY: Keep code DRY (Don’t Repeat Yourself) by avoiding duplicated code. Duplicated code is harder to maintain and increases the likelihood of introducing errors.
7. Test Your Code: Test your code regularly to ensure that it works as expected. Automated testing is particularly helpful for ensuring that code changes do not introduce new bugs.

Clean code is essential for any software development project. It ensures that the codebase remains manageable and efficient over time. By following best practices for writing clean code, developers can write code that is easy to understand, maintain, and test. Ultimately, writing clean code will lead to faster development, fewer errors, and more maintainable software.

What is Big Data?

Big data refers to the large, complex datasets that are generated by the digital age. It is characterized by its volume, velocity, and variety, as well as its high degree of variability and complexity.

In simple terms, big data is created by the vast amounts of digital information that are generated every day, such as social media posts, emails, digital images, videos, sensor data, and more. This data is then collected, stored, and analyzed to uncover insights and trends that can help businesses, governments, and individuals make better decisions.

The Four Vs of Big Data

Big data is often described in terms of the four Vs: volume, velocity, variety, and veracity.

Volume: Big data is generated in massive quantities, with datasets ranging from terabytes to petabytes or even exabytes.

Velocity: Big data is generated at a high velocity, with data coming in from multiple sources in real-time or near-real-time.

Variety: Big data is characterized by its variety, with data coming from a wide range of sources and in a variety of formats, including structured, semi-structured, and unstructured data.

Veracity: Big data is often described as being “noisy” or uncertain, with data quality varying widely and errors or inconsistencies common.

Why is Big Data Important?

The rise of big data has had a significant impact on a wide range of industries and fields. Here are just a few of the reasons why big data is so important:

1. Improved Decision-Making: Big data provides insights and trends that can help businesses and organizations make better decisions, improve operational efficiency, and reduce costs.
2. Personalization: Big data enables businesses to personalize products, services, and marketing messages to better meet the needs of individual customers.
3. Innovation: Big data provides a wealth of information that can be used to develop new products, services, and business models.
4. Healthcare: Big data is being used to improve healthcare outcomes by providing more personalized care, identifying and managing chronic diseases, and enabling early disease detection.
5. Fraud Detection: Big data analytics is helping to identify fraud and other financial crimes by analyzing large amounts of transactional data in real time.

The Challenges of Big Data

While big data has the potential to deliver tremendous benefits, it also presents several challenges. These include:

1. Data Security: Big data requires robust security measures to protect sensitive information from cyber threats and data breaches.
2. Data Quality: Big data is often “noisy” and requires sophisticated tools and algorithms to clean and process.
3. Talent Gap: There is a shortage of skilled data scientists and analysts to work with big data, which makes it difficult for organizations to make the most of this valuable resource.
4. Privacy: Big data raises serious privacy concerns, with many people worried about the collection and use of their personal data.

Conclusion

Big data has the potential to transform the way we live and work, from improving healthcare outcomes to making businesses more efficient and effective. While it presents several challenges, the benefits of big data are clear, and it is likely to play an increasingly important role in the digital age.

A quote from Pamela McCorduck states,

“Artificial Intelligence began with an ancient wish to forge the gods”.

For thousands of years, humans wanted to create something smarter than us, faster than us, something that can seemingly do things that we would never be able to do. But now suddenly we are starting to ask what happens if that wish comes true? what happens if we’re finally able to create these gods that roam about planet earth with all of us? Is that something we want? As the saying goes, Be careful what you wish for.

Today we see lots of different examples of AI. I’m sure many of you are familiar with this. Many of you are carrying this in your pocket. Siri, Cortana, Alexa, These are just the tip of the iceberg, there are many of these all over the place. And many of us are carrying them with us all the time. So AI is already very much intergraded into our lives and into how we work. But when we think about modern-day times AI had a little bit of a turbulent start. If we were to look back, the concept, idea of modern-day AI actually was created in 1956, At a conference in Dartmouth. This was the official creation of the term artificial intelligence, AI. This was the first time anybody has ever heard this phrase. After this term and concept was created at the conference there has been a lot of excitement in AI. The first golden years have been 1956 to 1974. During this time there has been a lot of investments, a lot of excitement to see AI happen. But unfortunately shortly after a lot of people have realized that they’re not making that much progress with AI. It has started to become expensive, it has started to take up a lot of time, maybe this AI stuff wasn’t going to happen at all. For a period of a couple of years, the conversation around AI was subsided a little bit. But shortly afterward a resurgence of AI happened.

from 1980 to 1987 rise of expert systems happened. An expert system was built on this logic of if then. If you input some kind of information then it would give some kind of output. Again a lot of organizations have put time, money, and resources into this. Yet again they have run into more roadblocks. The main problem in these systems was if you input some sort of a weird value, you would get some kind of jibberish out of the system. so this if-then expert system really only worked for standard data, standard information, stuff that was a little bit more predictable.

The major breakthrough in AI happened after the supercomputer named “Deep blue” beat the world chess champion, Gary Kasparov, in a chess match. And shortly after there was DARPA grand challenge that said “ We want an autonomous vehicle to drive through uncharted terrain by itself for at least 100 miles. And that challenge was too completed. And people clearly realized that something is now changed and we are moving in the right direction.

Since then AI field has been growing immensely and a lot of astonishing creations have made our lives so much easier. AI-powered solutions are used by dozens of companies and implemented in different fields, changing a lot of industries and reshaping the landscape of health, learning, daily living, and so on.

Some of the main achievements of artificial intelligence are,

1. Text Generation
2. AI Induced face recognition
3. advances in algorithm design
4. Drug discovery
5. speech recognition
6. lung cancer diagnosing

Security, coding, law, marketing, and many other fields are also transformed by AI. Technology evolves rapidly making intelligent systems even more intelligent and robust.

So, some things that were fantastic a few years ago are now a reality. Thus, thanks to AI our cars drive themselves, smart houses adjust electricity usage, programs analyze our DNA and diagnose illnesses. In other words, AI technologies and applications cover more in our lives, automate daily tasks and change the way we work. However, the best is yet to come.

That's it for today's blog folks. see you guys again with another interesting topic. Cheers!!

Venushki Thilakawardana

IT Undergraduate at University of Moratuwa Sri Lanka

https://twitter.com/VenushkiT

Venushki Thilakawardana - University of Moratuwa - Colombo District, Western, Sri Lanka | LinkedIn

Read my blog on :

Fundamentals of Cloud computing https://venushkithilakawardana.medium.com/fundamentals-of-cloud-computing-a568867ff35

What is machine learning

https://venushkithilakawardana.medium.com/what-is-machine-learning-e574c269ed16

Humans learn from their past experiences and machines follow instructions from humans. But what if humans can train the machines to learn from the past data and do what humans can do much faster? well, that’s machine learning.

The world is filled with data today. Generated not only by people but also by computers, phones, and other devices. this will only continue to grow in the years to come. Machine learning brings the promise of deriving meaning from all that data. Basically, it is tools and technology that you can utilize to answer questions with your data. Traditionally humans have analyzed data and adapted the changes in data patterns. However, as the volume of data surpasses the ability of humans to make sense of it and manually write those rules. We will turn increasingly to automated systems that can learn from the data and importantly, the changes in data to adapt to a shifting landscape. We see machine learning all around us in the products we use today. However, it isn't apparent that machine learning is behind it all. While things like tagging people inside of photos are clearly machine learning at play. It may not be immediately apparent that recommending a video to watch next is also powered by machine learning. Of cause, the biggest example of all is google search. Every time you use Google, you’re using a system that has many machine learning systems at its core.

Today machine learning’s immediate applications are already quite wide-ranging. including image recognition, fraud detection, and recommendation systems, as well as text and speech systems too. Basically, every company is pivoting to use machine learning in their products in some way. It’s rapidly becoming an expected feature just as we expect companies to have a website.

As I mentioned before the simple definition of machine learning is “answering questions using data”. in particular, we can split the definition into two parts. “Using data” and “answering questions”. These two pieces broadly outline the two sides of machine learning. Both of them are equally important. Using data is what we refer to as training. While answering questions refers to making predictions or inferences. Training refers to using our data to inform the creation and fine-tuning of a predictive model. This predictive model can then be used to serve up predictions on previously unseen data and answer those questions. As more data is gathered, the model can be improved over time and new predictive models can be deployed. As you may have noticed the key component of this entire process is data. Data is the key to unlock machine learning, just as much as machine learning is the key to unlocking that hidden insight in data.

This is just an overview of machine learning. Why it’s useful and some of its applications. Hope you enjoyed it and see you with another interesting topic.

Venushki Thilakawardana

IT undergraduate at university of Moratuwa Sri Lanka

https://twitter.com/VenushkiT

Venushki Thilakawardana - University of Moratuwa - Colombo District, Western, Sri Lanka | LinkedIn

Read my blog on :

Fundamentals of Cloud computing https://venushkithilakawardana.medium.com/fundamentals-of-cloud-computing-a568867ff35

Evolution of AI Technology https://venushkithilakawardana.medium.com/evolution-of-ai-technology-73639228a2b4

For all the non-techies out there the concept of cloud computing can be a little intimidating or puzzling.

So let’s see what’s all this fuss is about…

Simply put, cloud computing is techspeak that describes software and services that run through the internet rather than private servers and hard drives. This means instead of using your hardware and software for servers, data storage, networking, databases, etc, cloud computing allows you to use all these services as per requirement. The infrastructure needed to support the programs or applications that the consumer requires is owned and operated by a third party and the consumer pays only for the services they use. Since a lot of large and small businesses depend on their data, they have to spend fortunes on maintaining this data with strong tech support and storage space. so the cheaper option is operating on the cloud. This means whether it’s storage or bandwidth, you no longer have to pay for IT you’re not using.

There are four types of cloud models that you can use according to your business needs

There are three types of major services that cloud computing offers

1. software as a service (SaaS).

Software as a service also known as cloud application services is the most well-known of the cloud market. The software is hosted by the vendor or the service provider and made available to the client through their web browser. That means when clients are using SaaS there is no need to download and install applications on each computer.

2. Platform as a service (PaaS)

A lot of software developers, web developers, and businesses can benefit from PaaS. This provides a platform to develop, deliver and maintain applications including software support, management services, storage, networking, testing. Clients access this service via the internet.

3. Infrastructure as a service (IaaS)

IaaS gives clients access to infrastructures like network connections, bandwidth, virtual server space, load balancers, and IP addresses. Even though the client is responsible for managing their applications, data, etc., IaaS provides a high degree of control and flexibility to the client.

Now that you have a basic idea about cloud computing, let’s see the benefits of using cloud services.

Lower cost for the businesses that can’t afford the in-house infrastructure and backup support services.

A well-managed cloud platform is much more reliable than an in-house IT infrastructure because of the fewer downtime or service interruptions.

Assured backup and recovery of data.

Giving clients the ability to access the necessary information and communicate with each other whenever and wherever they want from any device that is connected.

greater efficiencies and economies of scale mean more money in your pocket in the long run.

That’s all for today's blog folks. See you again with another interesting topic. Cheers!!

Hope you enjoyed my very first blog.

Venushki Thilakawardana

IT undergraduate at University of Moratuwa Sri Lanka

Venushki Thilakawardana - University of Moratuwa - Colombo District, Western, Sri Lanka | LinkedIn

https://twitter.com/VenushkiT