As builders on Snowflake, we always faced challenges with understanding query patterns and optimizing them for performance and costs. In a world where the Data Cloud is being adopted by Snowflake partners, internal data engineers, data scientists, and even business teams, query observability and optimization becomes critical. We wondered, why should this remain a manual process for data teams, and whether AI agents could simplify this!
Presenting Snow-Wise, an AI agent that's aware of all activity happening inside the data warehouse, alerts the team in-charge in case of anomalies(in performance, latency, or cost), has reasoning/intelligence to suggest query optimizations, and even has access to Snowflake(as a tool) to verify its optimizations!
Data teams can interact with Snow-Wise via conversations today, but the architecture of the AI agent will very soon grow into further agentic patterns, that allow for automated schedules, human-in-the-loop, and self-learning behaviors.
In the demo video, you'd be able to see how a data engineer can simply ask in natural language which were the longest running queries in the last 3 days, and almost immediately, be presented with suggested optimizations and query logs that verify the veracity of the agent's suggestions
This AI agent was built on top of a multi-LLM setup(Snowflake Arctic and GPT-4), having access to the relevant Snowflake databases/tables like warehouse_metering_history and query_history.
GPT-4 is used for the purpose of orchestrating function calling, while Arctic is used for specifc Snowflake tasks like verifying query correctness.
The conversational user interface was built on Streamlit, which made it extremely simple to prototype and test the whole application.
We would've loved to use Snowflake Arctic for the entire LLM stack, but the lack of function calling in Cortex limited us. Hence, we had to switch to a multi-model setup.
- We learnt about Snowflake Arctic's capabiltiies, tried out different prompts and tool-calling patterns to identify how Arctic compared with other LLMs like GPT-4 and Llama-3+Groq. Eventually, we decided to stick to GPT-4+Arctic as our final stack.
- We also learnt the architecture paradigm of chaining together multiple tools and LLM calls to orchestrate user queries in a conversational flow, and how short-term history could be maintained across the whole chat session.
-
Snow-Wise can be extended to even suggest schema modelling improvements. By analyzing typical query patterns from query logs and the current schemas, Snow-wise could suggest ideas like denormalizing two tables into one.
-
This is something we're already building for our internal team. Snow-Wise will be given access to a Slack tool, and the agent would run regularly on a schedule to send anomalies and optimizations over Slack everyday.
-
If your Snowflake supports multiple warehouse-native solutions(like BI solutions), Snow-Wise could analyze all queries originating from them and share these as optimizations/suggestions to the relevant teams.