Agentic retrieval for 10× less

The Apache-2.0 core includes BM25, vector and hybrid search, plus SQL through DataFusion, single-node, on your object storage. Query DSL compatibility and Iceberg, Delta, and Hudi integration are included with Cloud and Enterprise, as shown in the table above. pip install infino, npm install @infino-ai/infino, or cargo add infino.

The Apache-2.0 retrieval engine runs single-node against S3, Azure Blob, or local disk. Cloud (beta) runs that engine for you and adds the managed API, catalog integrations, high availability, support, and SLAs.

Four things: storage, write tokens, read tokens, and returned bytes, plus an account minimum. Customers size the workload; Infino manages serving capacity.

A read token meters measured query work against a standard top-10 vector query. A simple keyword query uses fewer tokens; hybrid search or SQL over search results uses more. The same meter applies across cache states. Pinned tables assumed at higher volumes.

Pinned tables assumed at higher volumes. Heavier query shapes reach that point at lower QPS. Infino manages nodes, replicas, and cache placement while durable data remains in object storage.

A write token meters ingest work on the same unit as a read token. Volume follows the bytes written: a 1024-dimension vector is 4 KB. Production billing meters the work each write actually performed.

Each query mode applies the same workload inputs to the lowest-cost public-list-price deployment that a senior engineer could reasonably choose for production. Full-text compares Elastic and OpenSearch. Vector compares Qdrant, Pinecone, Elastic, and OpenSearch. SQL compares Postgres, ClickHouse, Snowflake, and Databricks. Hybrid includes the retrieval indexes or managed search services each option needs. Capacity, availability, and published vendor limits determine sizing; a row is omitted when no published configuration covers the workload. Hover a total for line items, and use the pricing methodology page for every sizing rule and source.

Elastic uses its published hosted data-out rate and shared allowance. OpenSearch, RDS, Qdrant Cloud, and Databricks use standard tiered AWS internet-egress rates (Qdrant Cloud and the Databricks lake run on AWS us-east-1). Pinecone includes 100 GB and then charges its published per-GB overage. ClickHouse uses its published public-internet rate. Snowflake uses its published AWS US East internet-egress rate.

Yes, through hydration. Infino adds a BM25 index and, when vectors are present, a vector index to ordinary Parquet. A later generic Parquet rewrite preserves the columns and requires the retrieval indexes to be rebuilt.

Cloud (beta) is multi-tenant serverless: the engine run by us, metered on storage, write tokens, read tokens, and returned bytes, scaling to zero between requests. Enterprise is a single-tenant deployment, on-premise or hosted for you, with multi-node queries, customer-managed keys, and dedicated resource isolation. It uses the same meters and is priced per agreement.

Through document size. A 1024-dimension 32-bit embedding is 4 KB, which usually dominates the document, so dimensions show up in the storage and write-token meters. Full-text and SQL run directly on text and columns.

Your data stays where it is, readable. Tables are standard Parquet files in your bucket, so DuckDB, pandas, and Spark keep opening them after the last Infino process shuts down.

Apache-2.0. The LICENSE file in the repo is the standard text, unmodified.