DataPup

Converts natural language questions into SQL queries by analyzing database schema and table relationships. The system ingests table metadata (column names, types, relationships) and uses an LLM to generate contextually appropriate SQL based on the user's intent, enabling non-SQL-fluent users to query databases through conversational prompts without manual query construction.

Abstracts database connectivity across multiple SQL and NoSQL engines (PostgreSQL, MySQL, MongoDB, etc.) through a unified client interface. Handles connection pooling, credential management, and schema introspection without requiring users to write database-specific connection code, exposing a consistent API regardless of underlying database type.

Executes generated SQL queries against the database and provides execution results back to the user, enabling iterative refinement. When a query fails or returns unexpected results, the system captures error messages and result metadata to feed back into the LLM for automatic query correction, creating a feedback loop that improves accuracy over multiple iterations.

Automatically discovers database schema structure including tables, columns, data types, primary keys, foreign keys, and indexes through database-native introspection queries. Builds an in-memory representation of table relationships and constraints that is passed to the LLM as context, enabling the model to understand how to join tables and respect referential integrity without explicit schema documentation.

Abstracts interactions with multiple LLM providers (OpenAI, Anthropic, local models, etc.) through a unified interface, handling provider-specific API differences, token counting, and prompt formatting. Implements domain-specific prompt engineering that structures schema context, query requirements, and error feedback in a format optimized for SQL generation, including few-shot examples and constraint specifications.

Validates generated SQL queries before execution to detect potentially dangerous operations (DELETE without WHERE, DROP TABLE, etc.) and enforces safety policies. Implements pattern matching and AST-based analysis to identify risky query structures, with configurable allowlists/denylists for tables and operations, preventing accidental data loss or unauthorized access.

Transforms raw database result sets into structured, displayable formats with metadata about column types, row counts, and data characteristics. Generates visualization hints (e.g., 'this is time-series data', 'this is categorical') that can be used by frontend clients to automatically select appropriate visualization types, and handles pagination/streaming for large result sets.

Maintains a history of executed queries, results, and user interactions to provide context for subsequent queries. Stores previous queries and their results in a structured format that can be referenced in follow-up natural language questions (e.g., 'show me the top 10 from the previous result'), enabling multi-turn conversations about data without re-executing queries or losing context.