Introduction
Northwind Health's clinical agents built with Microsoft Foundry retrieve information from drug formularies, clinical guidelines, and lab documentation to support clinical decision-making. The basic retrieval-augmented generation (RAG) pipeline you implemented using Foundry IQ works, but clinical staff report that agents sometimes return irrelevant documents or miss critical information when answering medication questions. When agents can't find the right knowledge, patient care decisions become slower and less accurate.
This module focuses on the knowledge layer of a clinical agent—the RAG pipeline that retrieves the right information when an agent needs formulary data or clinical guidelines. When an agent's reasoning chain determines it needs clinical knowledge to answer a medication question, the RAG pipeline decides which documents to surface and in what order. The pipeline you build here directly affects how accurately the agent supports clinical decision-making.
Note
This module covers retrieval and ranking before content preparation. You need to understand what search retrieves and how reranking works before you can optimize what goes into the index. Chunking and embedding strategies come last because they optimize the inputs to the retrieval pipeline you build first.
In this module, you design and implement production-grade RAG pipelines that address these limitations. You design hybrid search architectures that combine semantic similarity with exact keyword matching for clinical terminology. Semantic search is the foundational pattern this module builds on, retrieving documents based on meaning rather than keyword overlap using vector embeddings and dense retrieval. You implement multi-stage reranking pipelines that use successively more powerful models to identify the most relevant documents. You configure dynamic routing that directs queries to the appropriate knowledge sources based on clinical intent. You optimize chunking strategies and embedding models to match the structure of clinical content, including open-source embedding models from Hugging Face Transformers when Azure-hosted models don't match your domain vocabulary. By the end of this module, you build RAG pipelines that retrieve the right clinical knowledge consistently and efficiently.