OSI Field Mapping: every semantic layer, one spec

The name of a semantic object varies wildly across tools — cube, view, semantic_model, a raw SQL alias — but every product ultimately points at one physical table. OSI collapses that into two fields: dataset.name for the logical identifier your metrics reference, and dataset.source for the exact table binding downstream consumers hit.

OSI's fields[].expression.dialects[] is the one place multi-dialect SQL can live natively — MetricFlow, Cube and LookML each assume a single dialect. And fields[].dimension.is_time is the only cross-product flag any consumer can rely on for time dimensions: no more sniffing LookML's dimension_group, MetricFlow's type: time, and Snowflake's raw SQL to guess the same thing three different ways.

Snowflake Semantic Views and GoodData embed the aggregation inside the expression string (AS SUM(...), SELECT SUM({fact})). OSI moves it out to metrics[].aggregation so agents and BI tools can reason about the operation without parsing SQL. Derived metrics that reference other metrics are preserved verbatim through metrics[].expression.

This is the most inconsistent layer across the six products. MetricFlow and AtScale derive joins from entities / level bindings; Cube and LookML declare them inline; Snowflake and GoodData sit in between. OSI's relationships[] block gives every consumer the same four fields — name, from_dataset, to_dataset, foreign_key — plus explicit cardinality, which only Cube and LookML currently declare.

LookML's dimension_group with timeframes is the most complete; Snowflake leaves time entirely to SQL. OSI collapses this into a single time field with is_time: true and a granularity value — downstream tools that want day / week / month / quarter / year can generate them from the base field plus granularity metadata, so nothing is lost and every consumer can test one flag.

This is the layer OSI was built for. Snowflake shipped WITH SYNONYMS, AI_SQL_GENERATION and AI_VERIFIED_QUERIES in 2026 — and no other mainstream semantic layer has an equivalent yet. OSI standardizes those hints into ai_context on every field and metric, so an agent reading OSI can find synonyms, natural-language names and verified sample queries the same way whether the source stack is Snowflake, Cube or a home-grown YAML store. When downstream tools adopt OSI, the AI-grounding metadata travels with the metric — instead of being locked inside one vendor's SQL dialect.