// the find
kaarthik108/snowChat
Chat snowflake - Text to SQL
snowChat is a Streamlit app that lets you query a Snowflake database in plain English. It translates your questions into SQL using an LLM-backed agent, shows the results, and retains conversation context across turns. It's aimed at analysts or teams who want a quick chat interface over their Snowflake warehouse without writing SQL.
Supports multiple LLMs (GPT-4o, Claude Haiku, Gemini, Llama) so you're not locked to OpenAI. The self-healing SQL feature — where the agent catches errors and retries — is practically useful and often missing from simpler text-to-SQL demos. Schema context is ingested once and stored as embeddings in Supabase, which is the right approach for keeping prompts focused. The LangGraph agent architecture means tool calls are composable rather than hacked together in a single prompt.
Setup requires accounts in five separate services (Snowflake, Supabase, Cloudflare KV, OpenAI, AWS) just to run the demo — that's a lot of moving parts for what is ultimately a Streamlit prototype. The schema documents in /docs are hand-written markdown files you maintain manually; there's no automatic schema introspection, so your context drifts as the database evolves. Last commit was February 2025 and the repo shows no active maintenance, which matters if you're building on LangChain (it releases breaking changes frequently). Cloudflare KV is used for response caching but the cache invalidation strategy isn't documented, so stale results are a quiet footgun.