Wren

Converts natural language questions into executable SQL queries by parsing user intent through an LLM-powered semantic understanding layer, then mapping that intent to database schema metadata. The system maintains a semantic index of table and column definitions, allowing the LLM to reason about which database objects are relevant to the user's question before generating syntactically correct SQL that executes against the target database.

Enables querying across multiple heterogeneous databases (PostgreSQL, MySQL, Snowflake, BigQuery, etc.) through a unified natural language interface by maintaining separate semantic indexes for each database and routing queries to the appropriate backend based on table references detected in the translated SQL. The system handles cross-database join logic and result aggregation when queries span multiple sources.

Automatically generates human-readable documentation and semantic descriptions for database schemas by analyzing table names, column names, relationships, and data types, then enriching this metadata with LLM-generated summaries of what each table represents and how tables relate to each other. Users can also manually annotate schemas with business context, which is then incorporated into the semantic index to improve query translation accuracy.

Maintains conversation context across multiple turns, allowing users to ask follow-up questions that implicitly reference previous queries or results. The system tracks the conversation history, the last executed query, and result metadata, enabling it to resolve pronouns and relative references (e.g., 'show me the top 10' after a previous query) without requiring full re-specification. Context is managed through a sliding window of recent exchanges to keep LLM context manageable.

Automatically generates natural language explanations of query results, including summaries of what the data shows, identification of notable patterns or outliers, and business-relevant insights. The system analyzes result statistics (row counts, value distributions, aggregations) and uses LLM reasoning to surface actionable insights without requiring users to manually interpret raw data.

Enforces row-level and column-level access control by intercepting translated SQL queries and applying security policies before execution. The system logs all queries executed through the natural language interface, including the original natural language question, translated SQL, user identity, and results, enabling audit trails and compliance reporting. Access policies are defined at the database or table level and are applied transparently during query translation.

Caches previously executed queries and their results, allowing the system to return cached results for identical or semantically similar natural language questions without re-executing against the database. The cache is indexed by semantic similarity of the natural language input, not exact string matching, so variations of the same question can hit the cache. Cache invalidation is managed based on table update frequency and explicit refresh policies.

Allows users to define natural language questions as scheduled queries that execute on a recurring basis (daily, weekly, monthly) and automatically generate reports or notifications with results. The system translates the natural language question once, stores the resulting SQL, and executes it on schedule, then formats results into reports (PDF, email, dashboard) and distributes them to specified recipients.

Tracks the data lineage of query results by analyzing which tables and columns are referenced in the translated SQL, then provides impact analysis showing what downstream reports or dashboards depend on those tables. The system maintains a dependency graph of queries and their source tables, enabling users to understand data provenance and assess the impact of schema changes or data quality issues.

Automatically generates appropriate visualizations (charts, graphs, tables) for query results based on the structure of the data and the intent of the original natural language question. The system analyzes result columns (numeric, categorical, temporal) and the question context to recommend visualization types (bar chart, line graph, scatter plot, etc.), then renders interactive visualizations that users can customize.