Going Serverless with Search on Azure
Microsoft recently announced the availability of a serverless tier for Azure AI Search, bringing an elastic capacity to search workloads.
It now means that developers can send in burstable workloads where for example, the search volumes vary in different times of the day or where the amount of indexed data varies at the end of the month.
The problem with having a fixed cost to an Azure AI Search instance is that there can be costs incurred where it may not be necessary. For example, a search index may be provisioned first for a small data set for indexing, with the hope of utilising the quota later on but sometimes this may never happen due to project direction change. The costs of keeping an S tier Azure AI Search index will lead to a small project effectively costing more than it should, and that can feel disproportionate.
Enter Azure AI Search Serverless
As I hinted before the Serverless Developer tier is designed to do a few things that effectively quiet down the dials of an Azure AI Search instance and the activities around it, that cater to burstable workloads which are variable over time. These are the ways it does that according to Microsoft:
- Queries and indexing consume compute, measured in Compute Units per hour (CU/h).
- Active indexes consume compute based on their resource usage and how long they remain active.
- An index stays active for 10 minutes after its last query or indexing request before it goes inactive.
- Inactive indexes have no minimum or reserved compute charge. The compute usage for inactive indexes scales to zero. There's no minimum compute charge when an index is inactive.
- Storage is billed separately based on index size on disk and continues whether or not an index is in use.
- Agentic retrieval consumes compute for search queries and orchestration performed inside the Foundry IQ service.
Free and Basic Tiers still compete
Compared to the Free and Basic tiers, you do get a lot more indexes to work with at 10 000 , giving you more room to iterate across multiple iterations of indexes if you chose to do that as a Developer. With that quota given to you, you also exchange the favour by being limited to 10GB for your index size, which may still limit the ambition of each index, but much higher compared to free and Basic. As of the time of this writing, the Serverless tier is only available in 25 Azure Regions.
You will also have to sacrifice finer controls such as Partitions and Replicas (which are considered your Search Units in F, B, S tiers and higher) on the Serverless tier, they do not apply at all here. The architectural trade-off is that the way your AI Search instance behaves is handled in managed/serverless way, meaning you cannot control behaviour for when users are searching heavily or improve the query performance yourself.
Rather you have to become more strategic on how you send data to the index as one key tenet of using a Serverless tier. This can be done through smarter choices around index operations and requests:
- Indexing only new data or data that has changed to reduce the number of documents processed and ingested into the index.
- Having to skip heavy Image extraction up until you need it, this activity can still drive costs higher.
- Keeping index size small even when dealing with a higher quota on the amount of possible indexes.
- Being more mindful about Document enrichment pipelines with the chosen skills and skillsets by running them only on fields that truly require enrichment. Remember , you are being charged for each compute execution or consumption of compute as compared to having a fixed hourly charge in Basic for example.
- Not using index resets unnecessarily as these incur costs in pipelines that have heavy extraction and enrichment.

The choices and trade-offs are clear
The Serverless tier is for people wanting to make small and fast iterations over multiple projects and versions of indexes. If the workload is burstable, unpredictable and crucially small enough to not end up in run-away costs incursions that materialise through heavy enrichment enrichment pipelines, then a Serverless tier makes sense.
The Free and Basic tiers still make sense for people who want some predictability on storage quota, number of indexes and more control over index performance under load. The emergence of the Azure Serverless tier for AI Search should not change direction on those developers and projects