Revolutionizing Conversational Answering with RAGONITE

Conversational Question Answering (ConvQA) is a method that helps people find answers by asking questions in natural language. It works well with RDF Knowledge Graphs (KGs), which are just fancy ways to store lots of data in a way that makes it easy to search. The usual method involves taking a question and turning it into a type of query called SPARQL, which is designed for this kind of data. However, there are a few bumps in the road.

SPARQL can be pretty fragile, especially when questions get complicated. It isn’t great at dealing with abstract questions either. Instead of just sticking with SPARQL, we’re mixing things up with a new system that uses two approaches to get better results. We pull in data from a database that we create from the KG, using something called SQL Queries, and we also look for answers in text that describes the KG facts.

The cool part? If the first round of answers isn’t helpful, the system can go back for a second helping of information. This setup allows for an easy flow of information and results in clearer answers. We’ll explain all this in more detail and show how it works using BMW cars as an example.

#The Basics of Knowledge Graphs

Knowledge graphs store facts in a way that’s easy to understand. They use a simple structure often referred to as subject-predicate-object (SPO). This means that each fact is like a little sentence: something (subject) does something (predicate) to something else (object).

For example, you might have a fact that says "BMW X5 has a price of 50,000 EUR." In this case, BMW X5 is the subject, has is the predicate, and 50,000 EUR is the object. This structure allows people who manage the data to work without having to create complex rules like you would in traditional databases.

When someone wants to ask a question about the data in a KG, they usually use SPARQL. Think of SPARQL as a special language for asking questions that fit the graph format. However, with the rise of large language models (LLMs), more advanced ways of converting everyday language into SPARQL have emerged, simplifying the process significantly.

#The Two-Pronged Approach

While ConvQA systems make it easier to ask questions, they still face challenges. Often, parts of a question are left unsaid, making it tough for the system to understand what the user really wants. Even the best LLMs find it hard to translate complex questions into SPARQL queries.

To tackle this, we propose a two-pronged system. First, we run SQL queries over a database formed from the KG to address straightforward requests. Second, we use text searches to handle less clear questions.

The magic happens during the process where if the first results aren’t good enough, the system can automatically try again, pulling more information to give the user a clearer answer. This way, users can ask follow-up questions without losing track of the conversation.

#What is RAGONITE?

Enter RAGONITE, a clever system designed to tackle all these issues in a user-friendly way. RAGONITE stands for Retrieval Augmented Generation ON ITErative retrieval results. It has two main branches. One branch runs SQL queries and the other does text retrieval based on verbal explanations of the KG facts. It's like having two assistants, one who digs through the database and the other who reads the encyclopedia.

For instance, if someone asks, “What is the average acceleration time to 100 km/h for BMW Sport models?” RAGONITE processes the question, generates an SQL query, and looks for text passages that can provide context and details. If the answers aren’t satisfactory, it can go back to get more info.

#Making the Database

For RAGONITE to work its magic, it needs to create a database from the KG. This starts with converting the KG from one format (NTriples) to another that’s easier to work with (Turtle). Each fact is grouped by subject, and unique entities are identified to form tables.

We use the subject as a primary key, which allows us to track related information easily, like keeping your family tree organized. When new data comes in, it’s added to the right table, keeping everything neat and tidy.

#Verbalizing the Knowledge Graph

Some questions are tricky and require a little bit of common sense. For instance, if a user asks about the “innovative highlights in the BMW X7,” the system needs to interpret this better. To tackle this, RAGONITE verbalizes the KG facts into natural language passages.

This helps the LLM understand and respond to more abstract inquiries. The system uses simple rules to convert data into friendly sentences, ensuring that even the smallest details are captured.

#The Search Process

The retrieval process in RAGONITE involves looking for answers in two branches. If the first search doesn’t cut it, the system can repeat the process-like a dog fetching sticks until it gets the right one.

In terms of steps, the user inputs a question, RAGONITE creates an intent-explicit SQL query, and searches for relevant text. If the first attempt doesn’t provide a solid answer, it will suggest another round. The LLM uses this feedback to fine-tune its searches, making sure to look for the right information.

#Integrating Answers from Both Branches

As the two branches gather their results, they are brought together by the LLM to form a coherent answer. This way, the system doesn’t have to choose one method over the other. Instead, it blends the insights from SQL results and text passages to present a smooth response to the user.

This integration can even throw in citations to let users know where the information came from. It’s like having a personal guide who not only tells you the answer but also points to the exact page in the book where they found it.

#Open LLM Support

RAGONITE is designed to work with various LLMs, including ones that can be hosted locally. This makes it flexible and adaptable for different users' needs, especially for those who are concerned about data security. Local deployments of open LLMs like Llama-3 are also supported, providing greater access to the technology.

#Adding More Information

RAGONITE also allows extra text to be inserted into its backend system. For example, it can pull information from web documents, expanding its knowledge base beyond just the KG. This means that questions don’t have to be limited to the KG alone. If a user asks something more general, RAGONITE stands ready to provide those extra details.

#How RAGONITE Works in Practice

Imagine you’re using RAGONITE to find out if the BMW X1 is taller than a coupe. You type in your question, and the system immediately starts looking for answers. The first round might involve checking SQL results for the height of the X1, which might give partial information.

Then the system might switch gears and search through text records to find out how that compares to coupe models. Once it has gathered enough information, the final answer is generated and sent back to you, complete with references.

This design keeps the conversation flowing, allowing users to ask follow-up questions without losing track. It’s like chatting with a knowledgeable friend who is always eager to help.

#Results and Performance

When RAGONITE was tested, it performed impressively compared to its peers. The dual approach led to more correct answers than using either SQL or text retrieval alone. While both methods have their strengths, combining them leads to a more robust system.

The system showed particular strengths in tackling complex questions that others often struggle with. It handled abstract queries better by using verbalizations, bridging gaps that arise in conventional queries.

When it came to speed, RAGONITE was efficient as well. On average, the entire process of asking a question and getting an answer took just a few seconds. This quick turnaround time makes it practical for real-time interactions.

#The Future of RAGONITE

Looking ahead, the future of RAGONITE holds plenty of exciting possibilities. One aim is to enhance the system further, incorporating features that allow each part of RAGONITE to learn from its past mistakes and improve its responses.

Another goal is to perfect the integration of the various components, ensuring that they work seamlessly together. Fine-tuning specific parts of the system can also lead to even better performance over time.

#Conclusion

RAGONITE is an innovative system that combines traditional methods of data querying with modern conversational AI. By using both SQL and text retrieval, it addresses some of the biggest challenges in understanding user intent and delivering accurate answers.

It takes an intelligent approach to handling a diverse range of questions, proving itself as a valuable tool for anyone looking to dig deeper into the world of knowledge graphs and conversational AI. With such a system at your fingertips, asking questions about cars, or really any topic, becomes a lot less daunting and a lot more fun. So next time you have a burning question about BMW or anything else, RAGONITE might just be the buddy you need!