In the age of LLMs and AI-powered automation, crafting effective Smart Agents like the Virtuoso Support Assistant, OpenLink Data Twingler, OPML & RSS Reader, and ODBC & JDBC Connectivity Assistant involves a thoughtful blend of advanced technologies and user-centric design. This follow-up to my previous article delves into the technical architecture, design strategies, and steps involved in building these AI-driven assistants.
Conceptual Architecture
The image below depicts the interaction flow between a user issuing prompts to an Assistant and the actual implementation flow that occurs between the Virtuoso Server (a combined HTTP & Data Spaces Management system) and OpenAI Assistant Services.
1. Virtuoso Support Assistant: Knowledge-Driven Product Support
The Virtuoso Support Assistant provides expert-level product support by leveraging data spaces composed of knowledge bases, databases, and filesystem documents (e.g., product documentation). Here’s the component breakdown:
Predefined Query Templates & Retrieval-Augmented Generation (RAG): The assistant uses a combination of predefined query templates mapped to prompt patterns and RAG to search and retrieve relevant information in real-time. It draws from the data, knowledge graphs, and documentation harmonized in a Virtuoso-managed Data Space, ensuring seamless, context-aware responses.
Here’s an example of the Virtuoso Support Assistant JSON Configuration Document.
2. OpenLink Data Twingler: Simplifying Data Queries
The OpenLink Data Twingler is designed to allow users to execute complex SQL, SPARQL, and GraphQL queries within a ChatGPT or CustomGPT session. Here's how it was built:
Declarative Programming Focus: Data Twingler was built using a declarative approach, where the user describes the desired behavior (e.g., querying a database), and that's mapped to underlying stored procedures published as OpenAPI-compliant web services. The functionality for handling query statements is described in a YAML or JSON document. As with the Virtuoso Support Assistant, a JSON configuration file describes how prompts are mapped to query action invocations for a desired query language. This assistant offers the ability to query data from various ODBC- or JDBC-accessible database management systems, knowledge graphs, or file system documents.
Multiple Query Language Support: The assistant supports multiple query languages, including SQL, SPARQL, and GraphQL.
Natural Language Processing (NLP): The assistant’s natural language processing (NLP) capabilities translate user queries into executable queries in SQL, SPARQL, or GraphQL, enabling non-technical users to interact with complex data systems effortlessly.
Here’s an example of the Data Twingler Assistant JSON Configuration Document.
3. ODBC & JDBC Connectivity Assistant: Expert Product Support
The ODBC & JDBC Connectivity Assistant was created to provide expert-level support for database connectivity products. Here's how it works:
Predefined Query Templates (SQL): This assistant integrates with ODBC/JDBC data spaces and uses predefined SQL templates for querying product-related information. Like other assistants, it can access knowledge bases and documentation, enabling efficient troubleshooting.
Data Spaces Integration: The assistant interacts with databases, knowledge graphs, and file systems, ensuring the quality and accuracy of responses. This is crucial for addressing complex connectivity-related queries.
Here’s an example of the OpenLink ODBC & JDBC Assistant JSON Configuration Document.
4. OPML & RSS Reader: Accessing the Latest Content
The OPML & RSS Reader assistant is designed to streamline content consumption by retrieving and processing syndicated content from news sources and blogs. Here’s how it was built:
Feed Access: As with the other assistants, its behavior is described in a JSON configuration file where prompts are mapped to queries that retrieve content from feeds using their URLs. This capability extends to RSS, Atom, or OPML-formatted data representing articles or article collections. The assistant automatically responds to requests for the "latest edition" or similar terms, triggering change-sensitivity by leveraging HTTP-native cache invalidation when querying underlying data sources.
Information Retrieval: Like other assistants, it utilizes a combination of RAG and predefined query templates to retrieve relevant articles based on user prompts, also enabling article summaries for ease of use.
Here’s an example of the OPML & RSS Reader Assistant JSON Configuration Document.
The Technology Stack Behind the Assistants
All four assistants share a common foundation based on cutting-edge technologies:
OpenAI GPT-4: These assistants use GPT-4 for natural language understanding and response generation, allowing conversational interactions.
APIs and Connectors: Assistants are built using OpenAPI-compliant APIs that connect to external data spaces (databases, knowledge bases, knowledge graphs, and file systems), ensuring real-time retrieval and response generation.
Data Spaces Integration: In all cases, data spaces like databases, knowledge graphs, and filesystem documents play a critical role in improving response quality and ensuring contextually relevant answers.
Virtuoso Data Spaces Platform: A live Virtuoso instance provides the publicly accessible bridge for making this sophisticated functionality both demonstrable and usable in public and/or private settings. Its multi-protocol and user-controlled identity authentication functionality enables intelligent partitioning of data across LLM sessions.
Recapping How These Assistants Were Created: The Step-by-Step Process
Describe Assistant Behavior: Each assistant’s behavior is described in a JSON document, whether it’s querying complex databases, providing product support, or accessing the latest news.
Leverages Existing Open Standards: All data access is handled via open protocols like HTTP and OAuth, combined with declarative query languages such as SQL and SPARQL.
Fine-Tuning: Predefined query templates, combined with vectorized indexing of product-related content (in the case of Virtuoso and the ODBC/JDBC Support assistants), are key to ensuring high-quality responses.
Test and Optimize: Extensive testing is performed to ensure accuracy across various use cases, with a strong focus on response clarity and relevance.
Conclusion: Unlocking the Power of Declarative Programming and Descriptions
The Virtuoso Support Assistant, Data Twingler, ODBC & JDBC Connectivity Assistant, and OPML & RSS Reader represent the cutting edge of declarative programming and description-based interactions. By enabling users to focus on what they want rather than how it’s done, these Smart Agents empower both domain experts and everyday users.
These AI-powered tools demonstrate how LLMs are transforming automation and digital workflows, bridging the gap between data and actionable insights.
Feel free to ask questions using the comments section. We are in the midst of the most disruptive inflection to have hit the software industry in our life time!
Related:
The Evolution of Software-based Automation in the Age of Generative AI
Addressing DBMS Innovation Stagnation with Hyperlinks as Super Keys