Infino, founded by individuals with prior experience at OpenSearch, announced its emergence from stealth with a retrieval platform designed for AI agents. The core idea is to give an agent a single interface for querying structured and unstructured data, rather than distributing the task across a data warehouse, search engine, and vector database and then leaving the agent to combine the results over several rounds.
The company says agents differ from traditional users and applications in how they read. Instead of sending one query and waiting for the result, an agent asks small, concurrent questions to build a larger answer. Every result returned to it consumes part of the model’s context and may add time and cost to subsequent calls.
What does Infino offer?
Infino places retrieval functions directly inside SQL, allowing keyword search, semantic search, filters, counting, joins, aggregation, and sorting to be combined in a single query. According to CEO Ekechi Nwokah, the idea is not to use SQL itself, but to bring together query patterns that would typically require separate systems.
The company says a developer can enter a question in natural language and convert it into a single query that executes in fractions of a second. Infino has also added reasoning models aimed at the retrieval engine to handle repetitive tasks within search loops, such as query formulation, checking result quality, trying an alternative query, and verifying the answer, instead of always assigning these steps to more advanced, higher-cost models.
A single copy of data in Parquet format
The architecture stores data in Apache Parquet files on object storage, with search indexes embedded alongside the Parquet file footer. Infino says the core engine is open source and licensed under Apache-2.0 on GitHub, while the Infino Cloud service allows existing Parquet data to be used and made searchable without building the infrastructure in-house.
The data remains readable through tools that support Parquet, including Iceberg, Hudi, Spark, DuckDB, and other data warehouses. Parquet or JSON data can also be ingested and an agent connected to the Infino interface, without managing clusters, nodes, or distributed shards, according to the company’s description.
Why does this news matter?
If the idea succeeds in production workloads, a unified retrieval layer could reduce the number of ingestion paths, ETL jobs, schemas, and clients that an agent needs to call. More importantly from a security perspective, having a single copy of the data could enable access policies to be enforced centrally, including determining which rows and columns the agent can see and what is logged, rather than distributing permissions across MCP gateways and service connectors.
The platform targets platform and data engineers, security teams, machine learning and data science teams, and developers building agents that work with code, logs, continuous integration outputs, and tickets. Infino says its architecture is designed to be approximately ten times less expensive than traditional search and analytics architectures, while its published comparison shows a cost approximately 10.5 times lower than Elasticsearch and 23 times lower than OpenSearch for the workload it tested.
Limitations and open questions
These figures come from the company and are tied to a specific test workload, so they are not sufficient on their own to establish a general advantage over all Elasticsearch or OpenSearch applications. Reducing the number of systems also does not eliminate the need to verify retrieval quality, the limits of access policies, and index behavior as data changes or expands. The value of the approach remains linked to how well Parquet and object storage fit each team’s data and query patterns.