OKF vs Apache Ossie
Two open formats for giving AI agents the meaning of data. Open Knowledge Format (OKF) describes a model in markdown for agents and reviewers; Apache Ossie describes metrics, relationships and fields in structured YAML or JSON for tools to compute from. An October 2026 OWOX article concludes they are complementary and most teams will want both.
At a glance
| OKF | Apache Ossie | |
|---|---|---|
| Created by | Google Cloud’s Data Cloud team, announced 12 June 2026 | Started as Open Semantic Interchange (Snowflake-led); now an Apache Incubator project |
| Describes | The meaning of a model, in prose and structure | Datasets, fields, relationships, metrics and ai_context |
| Best reader | An AI agent, or a person reviewing a pull request | A BI tool, query engine or converter |
| Format | A directory of markdown files with YAML frontmatter, one concept per file; only type is required | One YAML or JSON document per model |
| Versioning | Optional okf_version in index.md; unknown versions are read on a best-effort basis | Required version string; some vendors treat an unsupported version as an error |
| Metrics | No metric schema; a metric is an example type value or an “Attested Computation” concept | A metrics list with an expression per SQL dialect |
| AI context | The prose in each file | Structured ai_context with instructions and synonyms |
| Provenance | OKF 0.2 adds optional sources, verification and lifecycle fields | No equivalent |
| Status | Spec at 0.2; v0.1 was announced as a starting point, not a finished standard | Draft 0.2.0.dev0 on the default branch; incubating |
| Licence | Apache 2.0 (Knowledge Catalog repository) | Apache 2.0 |
Neither format holds the SQL that builds the datasets; both describe the model, not the transformation.
What the OWOX conversion test found
OWOX converted the Apache Ossie TPC-DS example through its Model Canvas to OKF and back.
- Kept: 5 datasets, 4 relationships and 8 metrics, with the same counts after re-import
- Lost or changed: typed
ai_contextbecame plain text; 26dimensionblocks,unique_keysand the Salesforce and dbt extensions were lost - Formula change: a dataset prefix was dropped from one metric, so
SUM(store_sales.ss_ext_sales_price)becameSUM(ss_ext_sales_price) - Ossie to Ossie: this round trip also lost content; the file shrank from 672 to 519 lines
- A smaller shop model with three revenue metrics kept its formulas, but the canvas labelled a numeric revenue field as a string
When to use which
- OKF when an AI agent needs to understand tables and joins, or when people review changes in git
- Ossie when several tools must share the same metric definitions, or when typed structure matters
- Both: keep Ossie as the source and generate OKF from it, because the reverse direction drops structure
How it fits the vault
An SME with a Power BI semantic model can use Ossie to move definitions between tools (see Real-time business context for SMEs) and OKF to hand an agent readable context. This wiki itself follows the same markdown-and-frontmatter idea OKF formalises.
An OKF-style knowledge base in practice
AnomalyArmor generates per-table markdown knowledge bases from its monitoring signals, in a variation on OKF, and argues for them over RAG because retrieval flattens relationships such as foreign keys and lineage.
Source quality
The article is by OWOX’s Head of Product, and OWOX makes the Model Canvas used for the tests, though the article says neither format belongs to OWOX. It does not test LookML or other vendors’ converters.