Preprocessing Assets

ThothAI relies on retrieval assets that are prepared before normal querying.

Main Asset Types

• schema metadata
• comments and column descriptions
• evidence
• Gold SQL examples
• vector embeddings
• LSH-derived similarity artifacts

Why They Matter

These assets shape the context passed to the SQL generation agents.

They help the runtime:

• map business language to tables and columns
• recover relevant examples from the vector store
• reduce the schema for smaller generation models
• enforce evidence-aware testing before selecting a final query

Operational Consequence

If preprocessing is skipped or incomplete, the system can still run, but:

• context retrieval becomes weaker
• schema reduction becomes noisier
• SQL evaluation has less grounded evidence
• final selection quality drops