Infino launches a single retrieval layer for AI agents, stored in Parquet

An AI agent that answers one business question today calls a search engine for the keyword part, a vector database for the semantic part and a data warehouse for the numbers, then spends more model turns stitching the three replies together. Infino launched on Wednesday with a different shape: one retrieval layer that answers all four kinds of query from one copy of the data in Apache Parquet on object storage.

Infino says it was built by the engineers who created OpenSearch at Amazon, along with engineering leaders from LinkedIn, Google and Amazon. The core engine, infino-ai/infino on GitHub, is Apache 2.0 licensed. Infino Cloud is the paid hosted service that can point at existing Parquet files in S3, GCS or Azure Blob and make them searchable with no separate ingestion.

What the single-call pitch means

Infino CEO Ekechi Nwokah told The New Stack that agents are "the largest new consumer of data since the web browser" but are "being served with a fragmented stack built for a different era." A human writes one query and reads one answer, Nwokah said; an agent "issues many small questions to answer one large one, often at the same time, and every result it receives consumes context and can add cost and latency to subsequent model calls."

Infino's answer is to embed keyword, vector, filter and aggregate functions directly in SQL, so a natural-language question becomes one query that completes in milliseconds, by the company's own account. Nwokah said the important part is not SQL itself. It is that keyword search, semantic search, filtering, aggregation and grouping can happen together, rather than forcing the agent to orchestrate several systems and reason over the pieces.

How the Parquet layer is organised

Infino keeps a single copy of the data as valid Apache Parquet files on object storage, with its search indexes embedded next to the Parquet footer. Nwokah said any tool that reads Parquet can read the files with or without Infino. The open-source engine ships with that structure, and Infino Cloud adds the ability to attach to existing Parquet already sitting in a bucket. Table managers listed on the home page include Iceberg and Hudi.

Infino also found that agents spend a lot of time in retrieval loops, formulating a query, reading the result, deciding whether it is enough and trying something else. The company says it integrates smaller inference models beside the retrieval engine so those simpler steps run cheaper than through a frontier model, by its own account.

For a team already running Elasticsearch next to a Pinecone or Qdrant cluster next to a warehouse, the practical read is that Infino is pitching one ingest path and one client in the agent's tool list in place of three. Infino's own site also claims it is roughly ten times cheaper to run at scale. That cost figure is the company's own and has not been independently measured, so weigh it against your own workload before planning a migration.

Source

Primary source: Infino. Report: OpenSearch veterans launch Infino. Here's why it matters for agent builders. on The New Stack, by Adrian Bridgwater. GitHub repository: infino-ai/infino, Apache 2.0.