Invented by MANDAL; Sumangal, GUPTA; Vaishali, INTUIT INC.

Modern question-and-answer (QnA) platforms are becoming smarter, faster, and more helpful. This article explores a new patent application that aims to make QnA systems even better using artificial intelligence and improved search methods. If you’ve ever asked a chatbot for help or looked for answers on sites like Stack Overflow or Quora, you’ve seen these systems in action. Let’s break down what’s new, why it matters, and how this technology works.

Background and Market Context

People today expect fast, accurate answers online. QnA platforms have grown in popularity because they help users find information without waiting for a person to respond. Think of customer support chatbots on banking or shopping websites, or tech forums where you can ask for help with coding problems. These systems save businesses money and give users help at any hour.

But even with all this progress, there are problems. Many QnA platforms use a type of search called “vector search.” This search tries to match a user’s question with the closest question and answer already in the system. If you ask, “How do I reset my password?” the system looks for similar questions and shows you the best answer it can find.

Here’s where things get tricky: People don’t always ask questions in the same way. You might say, “I can’t log in—forgot password,” while someone else writes, “Password help needed.” These questions are different, but they mean the same thing. Basic vector searches can miss these connections, leading to poor results. Users get frustrated, and businesses lose trust.

This patent application tackles that problem head-on. It describes a way to make QnA platforms “smarter” by filling in the gaps. Using new AI tools, the system creates more ways to ask and answer questions, so users get better matches—even for questions no one has asked yet.

The market for these solutions is huge. Banks, online retailers, tech support centers, and educational sites all use QnA or chatbot systems. As more companies adopt these tools, the need for better search and smarter answers grows. This invention comes at a perfect time, meeting the demand for accuracy, speed, and reliability in digital help systems.

Scientific Rationale and Prior Art

Let’s talk about how things worked before and why change is needed. Traditional QnA systems store lots of question and answer pairs, taken from forums or written by experts. When a user asks something, the system converts the words into numbers—called “embeddings”—so it can compare the new question to old ones. This is the “vector search” mentioned earlier.

These embeddings are made using models like Word2Vec, GloVe, or FastText. These tools read through huge amounts of text and learn which words are similar. For example, “car” and “automobile” might be close together in this space, while “cat” and “engine” are far apart. The system turns sentences into sets of numbers (vectors), so it can quickly find which questions are most like the new one.

Vector search works well when users ask questions that are very close to what’s already in the database. But when someone asks a new or differently-worded question, the system often struggles. For example, if no one has asked, “Can I use my debit card internationally?” but there are many questions about using “credit cards abroad,” a basic model may not connect the dots. This leads to missed answers and unhappy users.

People have tried to fix this using better word models, ranking algorithms, and more. Some systems use machine learning to rank answers, or try to improve the data by cleaning it up. But the core problem stays: unless the exact question (or something very close) is already in the database, answers can be missed.

This is because the system only knows what it’s seen before. It can’t “imagine” new ways to ask the same thing, or create new question-answer pairs from the old ones. This is where large language models (LLMs) come in. LLMs, like GPT-4 or others, can read a question and answer, and then generate different—but related—questions and answers based on the same info. They can paraphrase, summarize, or even come up with new questions someone might ask about the same topic.

Before this invention, these two steps—vector search and LLM generation—were separate. Some research tried to use LLMs to help with QnA, but not in the way this patent describes. This technology brings together the power of LLMs and vector search in a new way, making it possible to “expand” the database automatically with new, AI-generated question-answer pairs. This fills in gaps and makes the system much more helpful.

Invention Description and Key Innovations

Now, let’s look at what this patent application claims and how it works. The invention is a system that combines traditional QnA data from the internet with new AI-generated question and answer pairs. Here’s how it happens:

First, the system gathers lots of QnA posts from well-known platforms like Stack Overflow, Quora, Reddit, or GitHub. A computer program called a “scraper” collects these posts. The system also filters out posts that are too long (for example, over 5,000 characters) to make sure the data is manageable and works well with AI models.

Next, each QnA post is sent to a large language model (LLM). The LLM is given a prompt—a set of instructions—to generate more question-answer pairs based on the original post. For example, if a post is about “how to transfer data from Kafka to S3,” the LLM might create related but new questions, like “What tools help with batch queries in S3?” or “How does Event Bus connect to a data lake?” Each new question also gets an answer, pulled from the original post or summarized by the LLM.

These new QnA pairs are then “embedded”—turned into vectors—just like the original posts. The system uses models like Word2Vec, GloVe, or FastText to do this. All these vectors (for both the original and AI-generated QnA pairs) are stored in a special database called a vector store.

When a user asks a question—maybe by typing into a chatbot—the system embeds the question and searches the vector store. It looks for the closest match using math methods like cosine similarity, Euclidean distance, or other ranking strategies. It can even use machine learning models to help sort the results. The best answer is then sent back to the user, often in real time.

This approach brings several key innovations:

1. Expanding the Knowledge Base Automatically: By using LLMs to create new QnA pairs from existing ones, the system doesn’t wait for new questions to be asked. It “imagines” what else users might want to know and adds those questions (with answers) to its brain. This makes it much more likely that any new user query can be matched to something helpful.

2. Mixing Human and AI Knowledge: The system blends real, user-generated content from QnA platforms with AI-created pairs. This means it benefits from real-world language and expert answers, but also covers topics or question forms humans haven’t added yet.

3. Smarter Filtering and Embedding: By filtering long or off-topic posts, and embedding everything with the latest models, the database stays clean and fast. Using models like FastText or GloVe helps capture meaning, not just keywords.

4. Flexible and Scalable Search: The vector store can handle millions of QnA pairs, both real and AI-generated. The search methods can be tuned: cosine similarity is fast for short queries, while neural network ranking can boost results for harder cases.

5. Real-World Performance Gains: The patent includes data showing how much better this system works. When only human-generated QnA pairs are used, the correct answer rate is about 78%. With only AI-generated pairs, it’s about 90%. When both are combined, the rate jumps to over 95%. That’s a huge improvement for anyone relying on automated help or chatbots.

The invention also covers how to keep the vector database updated, how to monitor chatbot queries, and how to use different types of user devices (phones, tablets, computers) smoothly. It even allows for different types of embedding and search algorithms, so the system can be adapted to different industries or needs.

In short, this system makes digital QnA platforms much more useful. It fills in the blanks, connects related questions, and gives users better answers, faster.

Conclusion

QnA platforms are everywhere, from customer support chatbots to tech forums and online classrooms. But they haven’t always been smart enough to answer every question, especially when users ask things in new ways. This new patent application describes a clever solution: combining AI-generated QnA pairs with real user data, then using advanced search methods to find the best match for any user query.

The result is a system that’s more accurate and helpful than ever before. By using large language models to expand its knowledge automatically, filtering and embedding data in smart ways, and searching with powerful algorithms, the invention sets a new standard for digital help systems. Businesses can cut costs, users get faster answers, and everyone benefits from a more connected, intelligent world of online QnA.

This technology isn’t just a small upgrade—it’s a big leap forward for anyone who relies on QnA platforms, chatbots, or automated support. As AI continues to evolve, expect to see even more powerful and user-friendly systems built on ideas like these.

Click here https://ppubs.uspto.gov/pubwebapp/ and search 20250219969.