Vanna.AI

Converts natural language questions into executable SQL queries by embedding your database schema into the model's context. Uses a retrieval-augmented generation (RAG) pattern where schema metadata (table names, column definitions, relationships) is stored in a vector database and dynamically retrieved based on query intent, then passed to an LLM for SQL synthesis. The model learns from your specific schema structure rather than generic SQL patterns.

Provides a unified Python interface to multiple LLM providers (OpenAI, Anthropic, Ollama, custom models) with automatic fallback and provider selection logic. Routes queries to the configured LLM backend without requiring code changes when switching providers. Handles provider-specific prompt formatting, token limits, and response parsing transparently through an adapter pattern.

Captures successful query-to-SQL mappings from user interactions and uses them to fine-tune or improve the underlying model's performance on your schema. Implements a feedback loop where correct SQL generations are stored as training examples, then used to retrain embeddings or adjust model weights. Works through a logging layer that intercepts user queries and their corresponding SQL outputs.

Manages connections to your database (SQL Server, PostgreSQL, MySQL, Snowflake, etc.) and executes generated SQL queries with connection pooling, timeout handling, and error recovery. Abstracts database-specific connection parameters and dialect differences through a driver abstraction layer. Handles query execution results and formats them for downstream consumption (pandas DataFrames, JSON, etc.).

Validates generated SQL queries for syntax errors, schema violations, and logical issues before execution. Uses a validation layer that checks if referenced tables/columns exist in the schema, detects invalid joins, and identifies queries that would fail at runtime. Provides error messages and can attempt automatic correction or suggest fixes to the user.

Maintains conversation history and context across multiple query turns, allowing users to ask follow-up questions that reference previous queries or results. Implements a stateful conversation manager that tracks the current query context, previous SQL generations, and result sets. Uses this context to disambiguate follow-up questions (e.g., 'show me the top 5' after a previous query) without requiring full re-specification.

Allows you to add business context, descriptions, and relationships to your database schema (table descriptions, column meanings, business logic notes). This enriched metadata is embedded into the model's context during SQL generation, improving the LLM's understanding of what each table/column represents and how they relate. Stores metadata in a structured format and retrieves it during query generation.

Implements row-level and column-level access control to restrict which data users can query based on their role or permissions. Enforces these restrictions at the SQL generation layer by modifying generated queries to include WHERE clauses or column filters based on the user's access level. Integrates with your authentication system to determine user permissions.

Generates SQL queries from natural language AND provides explanations of what the query does in plain English. Uses the LLM to both generate the SQL and produce a human-readable explanation of the query logic, helping users understand and verify the generated SQL before execution. Enables transparency and debugging by showing the reasoning behind the SQL generation.