Implement AI and analytics capabilities in Azure Cosmos DB
At a glance
-
Level
-
Skill
This learning path builds the AI and analytics capabilities that distinguish Azure Cosmos DB for NoSQL from a general-purpose document store. You start with retrieval, configuring full-text policies and vector policies on a container, choosing a vector index type against its dimension and volume limits, and running each kind of search against real catalog data.
You then combine them. Hybrid search fuses keyword and semantic results with reciprocal rank fusion, and you learn to account for its cost separately across each stage, avoid the filter placement that quietly removes the results the semantic half exists to find, and measure relevance against a fixed evaluation set rather than by impression.
With retrieval working, you ground a language model on it. You size a retrieval unit, separate the text you embed from the text you send to the model, assemble a prompt with verifiable citations, and treat retrieved content as evidence rather than instruction, which is the trust boundary that keeps a retrieved document from rewriting the system prompt.
The path closes with the two capabilities that make an application durable and measurable. You design an agent memory store that separates short-lived conversation state from long-term facts a later session can find by meaning, and you mirror operational data into Microsoft Fabric OneLake so analysts can query it with T-SQL and Spark without consuming a request unit on the database that serves your users.
Prerequisites
- Completion of the Design and develop Azure Cosmos DB database solutions learning path, or equivalent experience building applications on Azure Cosmos DB for NoSQL
- Familiarity with vector embeddings and calling an embedding API such as Azure OpenAI
- Programming experience with Python, which is the primary language throughout this path
- Basic familiarity with Microsoft Fabric workspaces and SQL
Achievement Code
Would you like to request an achievement code?
Modules in this learning path
Add keyword and semantic retrieval to an Azure Cosmos DB for NoSQL container without adding a search service. Configure full-text and vector policies, rank results with BM25 and vector similarity, and keep embeddings aligned with the text they came from.
Combine keyword and semantic retrieval in one query with Reciprocal Rank Fusion, and then tune the result for relevance, request charge, and latency. Isolate the search workload from transactional traffic when retrieval and writes compete for the same container.
Turn retrieval into answers. Prepare content so each result answers something on its own, assemble prompts that cite their sources and refuse when the evidence runs out, and wire the stages into a pipeline that survives a follow-up question.
Give an agent something to remember. Separate short-term conversation state from long-term memory, distill raw turns into facts a later session can find by meaning, and set the retention, reconciliation, and erasure policies that keep a memory store trustworthy as it grows.
Answer analytical questions about operational data without touching the workload that serves it. Mirror an Azure Cosmos DB for NoSQL database continuously into Microsoft Fabric OneLake, verify the copy is trustworthy, and query it with T-SQL and Spark.