Open Semantic Interchange (OSI) is a vendor-neutral YAML specification for moving semantic model definitions between tools, now developed as Apache Ossie (incubating) after a 2026-07-10 rename. Atlan is a launch partner. Its context layer for AI supplies lineage, ownership, and policy classifications alongside each definition. The latest released spec is 0.1.1.
Many pages, including some vendor docs, still say OSI. This page covers the current version, what the file holds, and what to test first.
- The spec describes one semantic model per YAML document, built from datasets, relationships, and metrics.
- Spec 0.1.1 is the released version, and 0.2.0.dev0 is a draft that says “Schema is mutable.”
- One vendor’s guide says most third-party tools do not import the file directly, and converters can drop fields.
- A metric file says what “revenue” means. Atlan’s context layer for AI says where it comes from, who owns it, and which policy applies.
| Item | Current state |
|---|---|
| Former name | Open Semantic Interchange (OSI) |
| Current name | Apache Ossie (incubating), renamed 2026-07-10 |
| License | Apache 2.0 |
| Latest released spec | 0.1.1 (2025-12-11) |
| Spec in development | 0.2.0.dev0, draft, schema may change |
What is Open Semantic Interchange (OSI), and why is it now called Apache Ossie?
OSI is the former name of Apache Ossie (incubating), a vendor-neutral specification for exchanging semantic layer definitions between analytics, AI, and BI tools. The project entered the Apache Incubator, and the new name was announced on 2026-07-10.
According to the project’s own update post (2026-07-10), the new name avoids confusion with other projects that share the OSI acronym. The same post says the spec, community, and mission did not change: “If you’ve been building on Open Semantic Interchange, nothing breaks.” One data cloud vendor’s docs still show both names, with the OSI-named procedures marked deprecated in favor of the Ossie-named ones.
A founding organization announced the effort on 2025-09-23, and a 2026-01-27 update said the first spec version was available in an Apache 2 licensed repository. Atlan’s launch announcement predates the rename, so it uses the old name.
A portable definition matters for the reason a semantic model does: a tool that redefines a metric on its own produces its own answer.
What does an Ossie semantic model file contain?
An Ossie file is one YAML semantic model with a version, a name, and at least one dataset. According to the current draft spec on GitHub (0.2.0.dev0), the structure looks like this. The draft moves the model to the document root, so check the version your tool reads.
| Element | What it holds | Required? |
|---|---|---|
version |
The spec version the file targets | Yes |
name |
The model’s name | Yes |
datasets |
Sources, keys, fields, and dataset-scoped metrics | Yes |
description |
A plain-language description | No |
ai_context |
Context meant for agents | No |
relationships |
Links between datasets | No |
metrics |
Model-level metric definitions | No |
custom_extensions |
Vendor-specific additions | No |
In the draft, only OSSIE_SQL_2026 is a dialect whose semantics the spec defines. Other dialect values are opaque, which means a tool reading them has to know what they mean. One model fits in one document, with no bundle format.
The ai_context field is where the spec points at agents. Version 0.1.1 lists “Context for agents” among its additions, so a model can carry guidance written for an AI reader next to the definition itself. How much of that a given tool reads is a separate question, covered below.
What does a metric look like across dialects?
A metric can carry expressions in several SQL dialects, and a vendor extensibility framework holds details one tool needs and another does not. One revenue definition can then serve tools that speak different SQL. The file describes business logic on top of physical tables, where a data model describes storage.
Which version of the Ossie spec exists today, and is it final?
Version 0.1.1 is the released spec, and the 0.2.0.dev0 on the main branch is an unreleased draft whose schema may change. The spec file itself labels 0.2.0.dev0 “Schema is mutable; do not depend on this version in production.”
| Version | Status | What it covers |
|---|---|---|
| 0.1.1 | Released, version history dated 2025-12-11 | Datasets, relationships, fields, metrics, multi-dialect metric expressions, vendor extensibility, context for agents |
| 0.2.0.dev0 | Draft in development | Unreleased; schema may change before 0.2.0 |
| 1.0 | None in the spec history as of 2026-11-05 | Not applicable |
The repository lists no GitHub releases, and its only tag is a release candidate, osi-0.1.1-rc1, so “released” here means the version recorded in the spec’s own history. Incubation carries a caveat too: the project’s site says incubation status is “not necessarily a reflection of the completeness or stability of the code.”
Some search summaries describe an “OSI v1.0 release.” The spec history lists no such version.
Who governs Apache Ossie and who is behind it?
Apache Ossie is governed through the Apache Incubator with “no single controlling entity” and a formal discussion-and-vote process for spec changes. Committership is earned by contribution.
The project’s 2026-07-10 update says the coalition grew from 17 launch partners to more than 50 organizations, and the repository opened in November 2025. A founding organization led the launch, so the accurate phrasing is a standard launched with partners including Atlan, not created by one company. The Apache proposal record, last updated 2026-06-16, lists Atlan among the contributing organizations, with code under Apache 2.0.
In April 2026 the project described five working groups: Advanced Metrics and Expression Language, Composability, Catalog Integration, Ontology Representation, and Model Converters and Developer Tools. The proposal’s planned work includes a metrics language, more converters, Apache Polaris integration, and a semantic query standard. Those are plans.
What reads and writes Ossie YAML today?
Several tools export or import Ossie YAML, but one vendor’s guide says most tools do not read it directly.
The converters repository lists 22 reference converter directories. Its README does not say which direction each converter supports, and the main repository does not accept new converters. The project’s ecosystem page, which carries no date, lists several products that import and export Ossie YAML, one that uses it natively as a knowledge-graph source, and two that expose Ossie models over MCP or REST.
A data cloud vendor’s developer guide (2026-10-14) is blunter: “Most third-party tools do not import OSI directly today.” The guide says its converters targeted 0.2.0.dev0 and were not on PyPI, and that its content is not maintained on an ongoing basis. That vendor’s docs page for creating a semantic view from Ossie YAML carries a Preview Feature label and says “Currently supported: 0.1.1.” Its release notes list the OSI YAML feature among open preview features, dated June 2026.
Conversions can lose information, as these cited sources show.
| Direction or tool | What was dropped | Source and date |
|---|---|---|
| One BI tool’s export | ai_context and the composite key |
Developer guide, 2026-10-14 |
| One metrics-framework converter | Refused the composite key | Developer guide, 2026-10-14 |
| A text-to-SQL service in the preview | Read description, not ai_context |
Developer guide, 2026-10-14 |
| One AI analyst product’s export | Targets spec 0.1.1; docs list model filters, glossary terms, grain and primary keys, and relationship types as not exported | Vendor docs, 2026-10-19 |
The file settles design-time portability; runtime access is a separate problem. An agent that resolves a metric through an MCP-served semantic layer does not need to parse YAML, and MCP delivering business context is the pattern Atlan’s context layer for AI follows.
OSI limitations: what the spec leaves out for now
The current draft spec names three things it does not support yet and one packaging limit. In its own words, with “Future versions may add these”:
- A field expression that references another field.
- A metric that references another metric.
- A dataset whose source is another dataset.
It also allows one model per document, with no bundle format and no cross-model references.
A metric built from other metrics, such as revenue per active customer, therefore has to be written as one expression, not as layers. Teams already using MetricFlow will recognize the shape of the problem, and the comparison with a metrics layer shows where definitions usually live.
The spec defines the structure of a semantic model. Owners, lineage, certification, and access policy are not elements of its schema, so a team keeps them in the systems it already runs.
What is the difference between OSI, a semantic layer, and a context layer?
OSI is a portable file format for semantic definitions, a semantic layer serves those definitions to queries, and a context layer for AI supplies lineage, ownership, and policy classifications alongside them. Each answers a different question.
| Question | Portable semantic file (Ossie) | Semantic layer | Context layer for AI |
|---|---|---|---|
| What is it? | A YAML spec for semantic models | A running system that serves metrics | The layer that gives agents definitions plus the context around them |
| What does it answer? | “How is this metric written down?” | “What is the value of this metric?” | “What does this mean, where does it come from, and which policy applies?” |
| Who or what reads it? | Tools that import it | BI tools, notebooks, apps, agents | Agents and people, through MCP, SQL, or API |
| What does it define? | Structure of a semantic model | Query-time metric logic | Meaning, lineage, ownership, and policy together |
| Where does Atlan fit? | Launch partner | Connects to the semantic layers a team runs | Atlan’s context layer for AI delivers definitions with lineage, glossary terms, and governance classifications |
What an agent receives is the point: a definition, its lineage, its owner, and the policy classifications that apply. The context layer vs data catalog vs semantic layer comparison walks through the split, and context engineering vs semantic layer covers choosing what reaches the model. Ontology vs semantic layer, semantic layer vs data catalog, and data catalog vs context layer separate the neighboring terms. The what is a context layer explainer and the semantic layer hub give the wider map.
Brad Shimmin, Vice President and Practice Lead, Data Intelligence, Analytics, & Infrastructure at Futurum, wrote in 2025: “The OSI has emerged as a crucial political treaty in a data ecosystem fractured by complexity and inconsistency.” A treaty settles shared vocabulary; it does not run the queries.
When does an OSI file help, and when should you wait?
An Ossie file helps when two or more tools must share one metric definition, and waiting makes sense when your tools need layered metrics or production-grade stability.
It helps when you are moving one semantic model between a warehouse and a BI tool, when you want to hand an agent a portable definition with ai_context attached, or when you want to review definitions as files in version control.
Wait, or test first, when the schema may change before 0.2.0, when converters are young and lossy, and when your tools do not read the file directly. One vendor’s import supports 0.1.1 only. Whether you need a semantic layer at all comes first, and the split between a semantic layer for BI vs AI agents changes who benefits. Text-to-SQL vs a semantic layer frames the tradeoff for self-serve analytics.
What should you do first if you are evaluating Apache Ossie?
Start with one model, one target tool, and a diff of what survives the round trip.
- Pick one high-value semantic model, ideally one with a contested definition.
- Check which spec version your target tool reads, and write against that version.
- Export, import, and compare field by field. Record anything dropped, such as
ai_context, keys, or glossary terms. - Decide where ownership and lineage live outside the file.
The wider path sits in semantic layer for AI agents, point solution vs context layer, and how to implement an enterprise context layer, with context engineering behind it.
How Atlan approaches semantic context for AI agents
Atlan’s context layer for AI gives agents the definition of a metric together with where it comes from, who owns it, and which policy classifications apply. A portable file can say what “revenue” is. The context layer adds the owner and the source.
Atlan announced its participation as an OSI launch partner on 2025-09-23. Prukalpa Sankar, co-founder of Atlan, said then: “Atlan’s mission is to be the context layer for the AI-native enterprise.” The proposal record lists Atlan among the contributing organizations.
What Atlan delivers sits next to the definition. According to Atlan’s own page on MCP delivering business context, the MCP server exposes asset metadata, lineage, business glossary and certified terms, governance classifications, and semantic context including metric definitions. The enterprise context layer is the system those pieces belong to, and the business context layer carries business meaning. Semantic views as materialized context covers one route for definitions.
Definitions move between tools, and the context around them has to move with them, or an agent gets a number without its meaning. A knowledge graph links definitions to owners and sources, the idea behind a knowledge graph for AI agents, and context engineering vs prompt engineering covers what reaches the model.
What a portable definition still leaves to you
A shared file settles how a metric is written. It leaves open who owns it, where it came from, and which policy applies, and those answers decide whether an agent can trust the number. Spec 0.1.1 is usable for a pilot. The draft schema, young converters, and tools that do not read the file directly mean the test comes first. How quickly converters catch up to the draft is still open.
FAQs about Open Semantic Interchange
1. What is semantic interoperability?
Semantic interoperability means different tools agree on what a metric or entity means. When a warehouse, a BI tool, and an AI agent all read one definition of revenue, they start from the same logic, which removes one major source of disagreement. Open Semantic Interchange is one attempt to make that possible through a shared file format.
2. Is Open Semantic Interchange the same as Apache Ossie?
Yes. The project entered the Apache Incubator, and the new name, Apache Ossie (incubating), was announced on 2026-07-10 to avoid confusion with other projects that share the OSI acronym. According to the project, the spec, community, and mission did not change.
3. Who started Open Semantic Interchange?
A founding organization announced it on 2025-09-23 together with launch partners, including Atlan. No single company controls it today. The Apache Incubator project describes “no single controlling entity” and a formal discussion-and-vote process for spec changes.
4. Is OSI a finished v1.0 standard?
No. The spec history lists no 1.0 as of 2026-11-05. Spec 0.1.1 is the released version, dated 2025-12-11, and 0.2.0.dev0 is a draft whose schema may change. The repository lists no GitHub releases, and its only tag is a release candidate, osi-0.1.1-rc1.
5. Can you use OSI files in production today?
Treat it as a pilot. One vendor’s docs label its Ossie YAML import a Preview feature, the draft schema may change, and conversions can drop fields. Test a round trip on one model first.
6. How is OSI different from a semantic layer?
OSI is a file format that describes semantic definitions so tools can exchange them. A semantic layer is the running system that serves those definitions to queries.
Sources
- Open Semantic Interchange (OSI) Updates: Specification Now Live, New Working Group Members and More, Snowflake
- Apache Ossie (Incubating): The New Name for Open Semantic Interchange, Apache Ossie
- Open Semantic Interchange: Community Update and What’s Next, Apache Ossie
- core-spec/spec.md, apache/ossie, GitHub
- apache/ossie-converters README, GitHub
- OssieProposal, Apache Incubator wiki
- Ecosystem, Apache Ossie (incubating)
- Getting Started with Apache OSI: An Open Source of Truth for Metrics, Snowflake
- SYSTEM$CREATE_SEMANTIC_VIEW_FROM_OSSIE_YAML, Snowflake Documentation
- Open Semantic Interchange format, Actian Documentation
- Is Open Semantic Interchange The Key For AI to Deliver Value?, Futurum
