Morningstar Intelligence Engine with Aravind Kesiraju — Weaviate Podcast #111
The Evolution of Enterprise RAG Systems
The journey of implementing Retrieval-Augmented Generation (RAG) and Agentic Systems at enterprise scale is marked by continuous refinement. Morningstar’s approach began with basic document processing and evolved through systematic experimentation with embedding, chunking, and retrieval strategies. One of the most interesting of these breakthroughs has been Contextual Retrieval, introduced by Anthropic in 2024, and Doc2query published by Nogueira et al. in 2019. Additional techniques, such as incorporating data overlap between chunks significantly improved context preservation, lead to improved responses.
Aravind describes their process reingested their large-scale document collection multiple times, each iteration bringing improvements to their ingestion techniques. Aravind additionally discusses the use of re-rankers for improved response quality.
Architecting Data Pipelines for Scale
Building reliable data pipelines for enterprise AI systems requires careful consideration of infrastructure choices. Morningstar opted for AWS services, utilizing SNS topics and SQS queues, complemented by Celery for robust task queue management. This architecture provides the flexibility needed to handle varying traffic levels while maintaining system reliability.
A particularly interesting aspect of their implementation is the automated pipeline integration with their CMS, enabling scheduled content processing with a pub/sub mechanism. This automation is crucial for handling the substantial daily content volume that characterizes enterprise operations.
The Tool Marketplace Approach
One of the most innovative aspects of the Morningstar Intelligence Engine is its tool marketplace. This feature allows teams to publish and consume API-based AI capabilities across the organization, enabling seamless integration of custom tools.
One particular Agent system implements the ReAct agent architecture. In the context of the API-based Marketplace, this allows for “agents within agents” — a powerful concept where one agent can utilize another agent’s capabilities through the tool marketplace. Learn more about the Google Cloud Marketplace in Weaviate Podcast #95 with Dai Vu and Bob van Luijt!
Text-to-SQL: Practical Lessons
The discussion of text-to-SQL implementation revealed several practical insights. Morningstar has achieved ~80–83% accuracy with Zero-Shot SQL generation, with some key learnings about database design around column naming and materialized views.
Column naming significantly impacts LLM performance. For instance, when dealing with ratings (gold, silver, bronze), having separate ‘rating_id’ and ‘rating_text’ columns helps the LLM understand which field to use for calculations.
Second, strategic use of database views can simplify the LLM’s task by reducing the complexity of joins it needs to generate. This approach of “doing more in the database” has proved particularly effective for complex queries.
Check out the Self-Driving Database from Andy Pavlo and team!
Looking Ahead: AutoGen and Multi-Agent Systems
Aravind shared valuable insights about their exploration of AG2 from Google Deppmind, formerly Microsoft Autogen, for multi-agent systems. While AutoGen offers greater flexibility than traditional approaches, it presents challenges for user-facing applications due to increased latency from multiple tool interactions. Their current approach is to use it primarily for backend processes, where response time is less critical.
Conclusion
The conversation with Aravind highlighted many aspects of building advanced Agentic systems. I hope this will inspire your interest in checking out Weaviate Podcast #111 with Aravind Kesiraju!
