28 July 2024

Introducing Stoqs

  • Vinay Saji Mathew
  • Bharat Jayprakash
  • Golnaz Bahrami

Stoqs is going to be a familiar product to many. It's built like most large language models on the market today, but a lot of thought has gone into presenting financial information in a useful and accurate way. Behind the scenes, Stoqs operates as a multi-agent system. It is designed to interpret natural language queries and aggregate high-quality information from multiple trusted sources for precise answers. Unlike relying on general web crawling, Stoqs orchestrates language models as "agents" to concurrently retrieve information from our curated sources. Additionally, we contextualize previous and current interactions to generate pertinent information necessary to address user queries effectively.

Financial research heavily relies on trustworthy data sources. Yet, much of this information is unstructured. Stoqs bridges this gap by efficiently curating and consolidating financial data. Stoqs makes highly accurate financial information accessible, empowering users with the insights necessary for informed decision-making. Whether it be market trends, investment strategies, or performance metrics, Stoqs is your answer.

So, what can you ask Stoqs?

Maybe you want to know

Maybe even this, and that.

The User Experience

Chatbots are highly linear. Almost everything is contained within a single conversation, regardless of how it may branch. However, when conducting financial research, comparison of multiple sources is important. Stoqs creates parallel threads for similar information within the same conversation. You can intuitively switch between threads by tapping on the tab selector at the top. Given the dynamic nature of the market, these threads have lifespans, reducing clutter and discarding stale information. Stoqs also crafts visualizations for queries. It can be hard to generate a coherent visualization relying on language models alone. We use a combination of generative AI and deterministic solutions to come up with accurate visualizations – and a variety of them

A Modular Architecture: Collaborative AI

Sequential decisions made by a single agent often lead to bottlenecks, an explosion of tokens, hallucinations, and unintended consequences. To address these issues, we utilize an agentic network. In our testing, this agentic network performs exceptionally well with real-time data and is effective when data is augmented through this framework. Agents collaborate to make decisions, from determining the data requirements for a query to making the necessary API calls and executing them. Internal data dependencies are managed through collaboration and information dissemination. The modular structure allows for fine-tuning at each level of the process, giving us excellent control over the product from end to end. Current literature also indicates that this structure provides clear advantages in terms of adaptability and versatility.

Intent + Hybrid Retrieval

Once a query reaches our system, agents determine whether live data, static data, or a combination of historical context and up-to-date information is needed. After setting up the plan, retrieval agents gather the relevant data streams. Based on the intended user experience, the data is organized into answer modules and visualization modules to form the final response.

General Availability

Stoqs is currently in beta, and we welcome enthusiasts to provide feedback on our platform. If you would like to try out our product, please sign up for the waiting list on our beta page. We will reach out to you as soon as possible. We plan to make the platform generally available this fall.

Final Thoughts

A large part of this effort is student-led. The importance of connecting research with data access cannot be overstated. If you are interested in collaborating, please reach out to us at research@eikasialabs.com.

Acknowledgments

  • Dr. Soundar Kumara
  • Microsoft for Startups
  • Financial Modeling Preps