UNLOCKING ENTERPRISE AI

AISQL extends SQL with primitive AI operators: AI_COMPLETE for text generation, AI_FILTER for semantic filtering, AI_JOIN for semantic joins, AI_CLASSIFY for categorization, AI_AGG and AI_SUMMARIZE_AGG for semantic aggregations. These operators compose naturally with traditional SQL constructs, allowing users to blend relational and semantic reasoning.

They support multimodal input via a new FILE data type, enabling queries over images, audio, and documents. The system handles the complexities of LLM context windows for aggregation operators by using hierarchical aggregation strategies.

The core innovation of AISQL is its native integration of LLM capabilities directly into SQL. This allows users to write declarative queries that seamlessly blend traditional structured data operations with semantic reasoning over unstructured content. No longer do users need to export data, write custom scripts, or build complex external pipelines.

This integration simplifies complex analytical tasks, enabling product managers to filter support transcripts by sentiment, join with product catalogs semantically, classify issue severity, and summarize feedback, all within a single SQL query.

LLMs are at the heart of AISQL's semantic capabilities. They power functions like semantic filtering (AI_FILTER), classification (AI_CLASSIFY), and summarization (AI_SUMMARIZE_AGG).

The system leverages various LLMs, including open-weight models (Llama, Mistral) and partner endpoints (OpenAI, Anthropic, Meta). Advanced techniques like adaptive model cascades ensure efficient and accurate use of these powerful, but often expensive, models by routing queries to lighter proxy models or powerful oracle models based on confidence scores.