Overview
This guidance, developed in partnership with Microsoft, explains how to keep Azure AI Search costs in check while enabling AI-driven workloads on your Nasuni data.
Cost Awareness Best Practices
Azure AI Search service core feature pricing is billed in Scale Units (SUs) according to the Pricing page.
Choose the lightest SKU that meets today’s requirements
Basic and S1 tiers expose the full modern API surface (vectors, semantic ranker, agentic retrieval) while charging the lowest hourly rate per Search Unit (SU). Review the tier guidance before assuming you need S2 or S3. Also, review the service limits for each tier/SKU to ensure they meet your solution needs.
Scale units grow linearly
One replica (compute node) × one partition (storage) is one Search Unit (SU). Because that multiplication is linear, a two-by-two topology costs four times as much as the starter one-by-one, yet delivers little value if neither storage nor compute is a bottleneck. Add partitions only when index size or ingestion throughput requires it; add replicas when queries per second (QPS) of your solution requires them, when there are very complex queries that are throttling your service, or when high availability is required. Service limits show exactly how many documents, indexes, index size, and many other factors each combination supports. Review capacity management before setting up your service for additional considerations.
Use a capacity-planning worksheet before provisioning
Index 1-5% of representative content for initial testing and capacity/costs projections, including OCR, embeddings, or other skills you plan to use in your setup (if any), then extrapolate the index-to-source ratio (typically a fraction of original size), the index-time throughput you observed and costs. Refer to capacity management.
Budget the indexing and query pipelines, and additional network features
If you are using AI enrichment, image extraction, computer vision calls, embedding requests, custom skill application/function, and any other transformation you perform or external service to AI Search you call, when using skillsets, they run on separate meters. Review each skill you plan on using for pricing reference.
At query time, review the pricing page to see which premium features you are using that incur extra costs, such as the semantic ranker. Also, if you are using vectorizers, such as the Azure OpenAI vectorizer, review its pricing conditions.
Track live costs
Create a Cost Management budget alert for the resource group that contains your search service and any enrichment-related storage resources.
Key Documentation and Cost-Estimation Tools
Azure AI Search Pricing page – per-Scale Unit rates and separate meters for Semantic Ranker and other add-ons.
Azure Pricing Calculator – model replicas, partitions, and geographic prices in one place. Calculate pricing for all the additional services if used during the indexing or query pipeline.
Capacity planning article – formulas and checklists for pilot sizing.
Monitor queries & storage – built-in metrics for QPS, throttling, and index size.
Indexing Strategies
During pipeline design:
Consider caching skill outputs and using knowledge store so you pay for vision extraction or embeddings only once.
Use incremental indexing against SQL, Cosmos DB, so only new and changed rows are crawled. Blob storage and other indexes have automatic incremental indexing. Review each connector documentation for details.
Keep vector payloads compact. For vector search, consider vector compression best practices.
Additional Insights and Resources
Semantic Ranker, integrated query vectorizers, and Agentic Retrieval all introduce extra meters; they are opt-in at query time, so gate them behind an application flag for cost-sensitive workloads.
Hybrid and vector compression blogs provide relevance/cost trade-offs proven in Microsoft research. Hybrid retrieval blog and Vector compression blog.
Azure Monitor can stream metrics to Log Analytics; build dashboards that overlay QPS, latency, and cost so you know when to add or remove replicas.