Building an agentic Text-to-SQL pipeline for financial data. Is this architecture good?
retained text
Hey everyone, I'm building a Text-to-SQL chatbot so users can ask natural language questions against a company's financial database. I want to make sure this workflow is robust and not over-engineered. Here is my current step-by-step pipeline: 1. **Classifier:** A fast, cheap LLM rewrites the user's question and extracts entities. 2. **Cache:** Checks if this exact question was answered before. If yes, return the answer. If no, proceed. 3. **Schema Lookup:** Checks if it needs more context, then pulls the exact table structures/DDL needed. 4. **Business Rules:** Injects custom company definitions (e.g., how "premium tier" is calculated) into the prompt. 5. **SQL Generation:** A heavy LLM (like GPT-4o) takes all that context and writes the SQL query. 6. **Security Check:** A Python script scans the SQL for malicious injections or drops. 7. **Execution & Auto-Fix:** Runs the query in MySQL. If MySQL throws a syntax error, the agent catches the error and asks the LLM to fix it automatically. 8. **Final Grader:** A fast LLM checks the final data to ensure it actually answers the user's original question. 9. **Return:** Send the data to the user (or kick to a human reviewer if it failed). **My questions for the community:** * Is this a solid architecture for financial data, or is there a better/standard way to do this? * Is relying on a Python script for security (Step 6) a bad idea? * Will running this many LLM nodes sequentially cause massive latency issues? * Are there better architectures out there for this specific use case? Would love any feedback or advice on how to optimize this!
document fetched 2026-09-29 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wshhxr/building_an_agentic_texttosql_pipeline_for