Standard

Open Semantic Interchange (OSI)

A vendor-neutral, Apache-2.0-licensed standard (v0.1, January 2026, co-founded by Snowflake) for exchanging semantic models — metrics, dimensions, organizational relationships — across BI and AI platforms.

3 connections1 resources1 post

Summary

What it is

A vendor-neutral, Apache-2.0-licensed standard (v0.1, January 2026, co-founded by Snowflake) for exchanging semantic models — metrics, dimensions, organizational relationships — across BI and AI platforms.

Where it fits

The logical-layer analog of what Iceberg did for physical tables: portable business meaning. For AI agents querying lakehouse data on S3, OSI is what lets the model inherit governed definitions instead of inferring them — removing a hallucination class that comes from un-governed retrieval.

Misconceptions / Traps
  • v0.1 is a founding spec, not an adopted ecosystem — coverage currently traces mainly through Snowflake's own materials; watch for independent adoption before treating it as settled.
  • OSI does not replace data contracts (schema/quality/SLA); it carries the meaning layer above them.
Key Connections
  • scoped_to Lakehouse / LLM-Assisted Data Systems — the semantic layer over lake data
  • enables RAG over Structured Data — governed definitions ground agent queries

Definition

What it is

A vendor-neutral open standard (v0.1 finalized January 2026, Apache 2.0 license) for exchanging **semantic models** — metrics, dimensions, and organizational relationships — across BI and AI platforms. Co-founded by Snowflake. Where Apache Iceberg unified the *physical* storage of tables, OSI aims to unify the *logical business meaning* of the data those tables hold.

Why it exists

An AI agent querying an S3 bucket can read the bytes but not the business context — what "net revenue" means, which dimension is canonical, how entities relate. Every BI tool historically encoded that semantic layer in a proprietary format, so meaning was lost at each platform boundary. OSI's bet: a portable, machine-readable semantic model lets an LLM querying lakehouse data understand it the way a human analyst does — removing a class of hallucination risk that comes from un-governed retrieval, not from the model.

Primary use cases

Portable semantic layers across BI/AI platforms, grounding AI agents in governed business definitions before they query object-storage-resident data, semantic-model interchange between lakehouse vendors.

Recent developments

Latest signals
  • v0.1 finalized January 2026, released under Apache 2.0. The founding pitch is explicitly the Iceberg analogy: standardize the semantic layer the way the table format standardized storage, so metrics and dimensions survive the trip between platforms — and so agents inherit governed meaning rather than inferring it. Per Snowflake — open and interoperable lens.

Connections3

Outbound3

Resources1

Featured in