Pablo Castro is a corporate vice president and distinguished engineer at Microsoft, where he leads the CoreAI division’s AI Knowledge team across Foundry IQ, Azure AI Search, and Azure Content Understanding. His work equips AI agents to find, interpret, and securely use organizational information, drawing on a career that spans database architecture, enterprise search, and machine learning.
From databases to AI knowledge systems
Before Microsoft, Castro worked on distributed inference systems for credit scoring and risk analysis, alongside collaboration software. He subsequently became a technical lead on SQL Server, contributing to SQL-CLR integration, extensible type systems, the Tabular Data Stream protocol, and ADO.NET. He worked on the ADO.NET Entity Framework, led ADO.NET Data Services, and later directed engineering and data science for Azure Applied AI before overseeing engineering for Azure Synapse/SQL and Azure Cognitive Search.
His 2018 introduction of Cognitive Search combined Azure Search with AI-powered content extraction. By 2024, his focus had expanded to production retrieval-augmented generation: managing changing datasets, growing query volumes, complex document formats, and the tradeoffs among search quality, latency, and cost.
In Foundry IQ, Castro advances a reusable knowledge layer that connects agents to enterprise and public information through serverless retrieval and permission-aware security. Knowledge bases can combine structured datasets, unstructured documents, and web content while exposing lower-level indexing controls when developers need them.
Intrinsic, extrinsic, and learned knowledge: Castro separates what models acquire during training, what agents retrieve from external sources, and what organizations learn from agents’ actual behavior. His framework for AI and knowledge makes all three necessary components of useful enterprise agents.
Hybrid lexical and vector retrieval: Effective retrieval combines keyword matching and embedding similarity, then uses cross-encoder reranking to evaluate promising documents against the original query. Metadata filters further narrow the search space, improving relevance without applying expensive ranking models to an entire index.
Vector quantization with explicit tradeoffs: Compressing vectors increases storage density and can accelerate search; retaining full-precision representations permits more accurate reranking afterward. Castro’s production retrieval architecture also incorporates continuously refreshed ingestion, document chunking, and incremental indexing.
Agentic retrieval and measured learning loops: Multi-step retrieval can improve evidence recall and answer completeness for difficult questions, provided its additional latency and token costs are justified. Castro also applies execution traces to generate evaluations, establish baselines, test alternative agent configurations, and adopt changes only when measured performance improves.
Agents need more than knowledge stored in model weights: they need access to organizational context and a way to turn experience into better instructions, tools, and skills.
Production RAG needs more than vector similarity: hybrid retrieval, careful candidate budgets, stronger ranking, compact indexes, and ingestion that keeps pace with changing data.
