Backed By Databricks
Enterprise-grade agent platform for analytics and data workflows.
Databricks Mosaic is a production-ready agent framework enabling developers to build, deploy, and evaluate AI agents on enterprise data using RAG (Retrieval-Augmented Generation). Features include governance, quality evaluation via MLflow, telemetry/observability, security guardrails, and cross-model support. Widely employed in data analytics use cases across industries.
Enterprise analytics agents
Automated report generation
Data-driven decision agents
Production-quality RAG applications
Safety Filtering (via Meta Llama 3)
Monitor and audit request/response data being sent to model APIs
Monitor operational usage on endpoints
Unified Access to any model (Foundation, External, Custom)
Permission and Rate Limiting (User and Group-based)
Payload Logging to Inference Tables (for audit/monitoring)
Usage Tracking to System Tables (for cost and operational monitoring)
Traffic Routing (Load balancing/Fallbacks)
Real user experiences from across different platforms
Mosaic AI Gateway allows us to securely consume any LLMs, be they OpenAI or other models hosted on Databricks, while ensuring LLM traffic is properly governed and tracked.
Harisyam Manda
Recent
, compliance (PII, safety), and cost controls over a mix of internally hosted and external LLMs (e.g., OpenAI).
akehouse and Unity Catalog for data and AI governance.
external LLM and don't require enterprise-level governance.
Users requiring a free or fixed-cost solution.
Provides centralized governance, full observability, and built-in safeguards to securely and compliantly manage all GenAI model consumption (including external models) at scale without needing to rewrite application code for model switching or integration.
AI Guardrails and Fallbacks are not supported on custom model serving endpoints (Only rate limiting and usage tracking are supported for non-route-optimized custom models). AI Guardrails and Fallbacks are not supported for Mosaic AI Agents (yet). When AI Guardrails are used, the request batch size cannot exceed 16. AI Guardrails are not applied to requests and intermediate responses of function calling (only the final output response). Text-to-image workloads are not supported. Any request/response over 1 MB will not be logged in Inference Tables. The AI Guardrails moderation service is limited to regions that support Foundation Model APIs pay-per-token.