Goodbye Lucene.
Hello Parquet.
Infino replaces Lucene with Parquet files on object storage at trillion scale. Get 10× cost reduction on Elasticsearch or OpenSearch with a single click.
Save 10× on
Observability
Logs, metrics and traces. Your dashboards and alerts keep working in OpenSearch Dashboards.
RAG
Agents answer from your data over MCP or A2A, with the rows behind every answer.
Analytics
SQL with joins and windows over the same Parquet, with search inside the query.
One command to migrate.
Zero application changes.
- Your Elasticsearch and OpenSearch clients
- Your dashboards, from Kibana or OpenSearch Dashboards
- Index names, mappings and aliases
The fastest keyword search.
Top-10 keyword search over 1 billion web pages. benchmarks →
The fastest vector search.
p99 latency at 95% recall over 1M vectors, on the Open Vector Database Benchmark. the board →
The fastest SQL.
vCPU-seconds per query over 100M rows, lower is better. Infino on ClickBench, Snowflake and Databricks on CostBench. how it is measured →
Store, search, and analyze trillions of documents in object storage.
Each index is a superfile: a valid Parquet file with the data and its BM25 and vector indexes inside. Search, vectors, SQL and your agents all read that one copy in your bucket, and any Parquet reader opens it.
logs-app.parquet
rows
BM25 index
vector index
in your bucket
With an automatic context layer for agents.
Infino fuses small language models with retrieval to accelerate context for agents.
Empower all your teams to write agents with no ETL and no hand-written semantic model.
More answers exactly right.
Databricks' OfficeQA Pro V2: 90 questions over 1,435 U.S. Treasury documents, share answered exactly right. Infino answers at $0.63 a question, against $4.00 for Databricks Genie.
What you get. What you give up.
Infino replaces Lucene inside OpenSearch. Here are the tradeoffs.
What you get
- Faster queriesTop-10 keyword search in 78 ms against Lucene’s 536 ms. A sum and average over 10M documents in 1.6 ms against 89 ms.
- Lower costOne copy in object storage instead of replicas on disks. About 10× less than Elastic.
- Less maintenanceStateless nodes. A node with an empty disk brings every index back from your bucket, and compaction runs on its own.
- Full SQLJoins, aggregates and window functions over your indices, beside the Query DSL.
- Parquet compatibilityEvery index is Parquet in your bucket. Any Parquet reader opens it.
- Vector, hybrid, and graph searchkNN and hybrid queries on the same tables as your text.
- Agent automationAgents query your data over MCP, and anyone can ask in plain English.
- More accurate answers73.3% of OfficeQA Pro answered exactly right with Supermodel, against 55.6% for the frontier model alone.
What you give up, for now
- Some query typesfuzzy, function_score, geo and suggest, and fuzzy or proximity syntax inside query_string.
- Custom analyzersText is indexed with the standard analyzer only.
- A few slower shapesSorted searches are still faster on Lucene.
- Some write pathsUpdate scripts, sliced scroll, reindex and update by query.
- Mixed searchesOne search across Infino-backed and Lucene indices together.
Point Infino at your cluster.
or run it yourself: pip install infino · Apache-2.0 · quickstart →