The system uses Claude 3.5 Sonnet running inside Snowflake's secure perimeter for handling complex queries across both structured and unstructured data. This allows maintaining data governance while leveraging LLM capabilities.
Their search stack is using a hybrid approach combining vector and lexical search with semantic reranking. The benchmarks show a 12% improvement over OpenAI embedding models (measured on NDCG@10) using their Arctic-Embed-L-v2.0 model.
For structured data analysis, they're using semantic modeling rather than just pattern matching for Text-2-SQL generation. This helps handle complex multi-table environments and prevents common issues like JOIN hallucinations and double counting.
The quickstart [1] shows how to implement this in practice. Key components:
Agents that can plan multi-step queries Semantic SQL generation for structured data Hybrid search for unstructured content
There's a demo video [2] showing a practical insurance claims scenario that demonstrates:
How it handles complex document collections Real-time semantic understanding of data structures Multi-step query planning and execution Integration between structured DB queries and unstructured document search
Would love to hear from anyone who's worked on similar orchestration challenges across structured and unstructured data securely, and gain meaningful insights.What challenges have you encountered with mixing structured and unstructured data processing?
[1] https://quickstarts.snowflake.com/guide/getting_started_with... [2] https://www.youtube.com/watch?v=SDqyB-yWaAE