Name: metabase-representation-format
Author: metabase

Understands the Metabase Representation Format — a YAML-based serialization format for Metabase content (collections, cards, dashboards, documents, segments, measures, snippets, transforms). Use when the user needs to create, edit, understand, or validate Metabase representation YAML files, or when

Metabase Representation Format Metabase represents user created content as a tree of YAML files. Each file is one entity (a collection, card, dashboard, etc.). The format is portable across Metabase instances: numeric database IDs are replaced with human readable names and entity IDs. The format is defined by a spec that the @metabase/representations npm package ships (see Reading the spec ). The same package ships a CLI ( npx @metabase/representations validate schema ) that validates a tree of YAML files against the format. Entities The format defines 14 entity types. Entity SerDes Model Description Collection Collection Folder like container for organizing content. Hierarchy via parent id . Namespaces: null (main), "snippets" , "transforms" , "data actions" , "data apps" . Card Card Question, model, or metric. Holds an MBQL or native dataset query . Display types: table, bar, line, pie, scalar, etc. Card types: "question" , "model" , "metric" . Dashboard Dashboard Grid layout (24 columns) of cards with filter parameters and optional tabs. Contains dashcards array for card placement and parameters array for filter controls. Document Document Rich text page using ProseMirror AST. Can embed cards via cardEmbed nodes and link to entities via smartLink nodes. Segment Segment Saved filter definition scoped to a table. Definition is a pMBQL query with a single stage containing only source table and filters . Measure Measure Saved aggregation definition scoped to a table. Definition is a pMBQL query with a single stage containing only source table and exactly one aggregation . Snippet NativeQuerySnippet Reusable SQL fragment referenced in native queries via {{snippet: Name}} . Transform Transform Materializes query or Python script results into a database table. Source is either MBQL/native query or Python script. TransformTag TransformTag Label for categorizing transforms. Built in types: "hourly" , "daily" , "weekly" , "monthly" , or null for custom. TransformJob TransformJob Scheduled job (cron) that executes transforms matching specific tags. TransformTest TransformTest Runs a transform against test inputs and checks its output against expectations . Stored next to its transform: collections/transforms/{transform slug}/{test slug}.yaml . PythonLibrary PythonLibrary Shared Python source file available to Python based transforms. Action Action Writes to a database. type is "query" (parameterized native SQL) or "implicit" (one row of the table its implicit entry names); model id is deprecated. Without a model id it lives in the data actions namespace (its root or a data actions collection) or in a data app's collection, set by collection id ; data apps run only those. Exactly one of the nested query and implicit arrays holds the definition its type names; the other is [] . Data App DataApp A custom React app that runs inside Metabase, in data apps/<slug / : a data app.yaml manifest (with serdes/meta [{model: DataApp}] , and slug equal to the directory name), its built bundle beside it, and a root Collection of the data apps namespace under collections/data apps/ holding the Cards and Actions it runs. Ownership and hierarchy Critical — folder layout is decorative. Where an entity lands in Metabase is decided entirely by its fields, not by where its YAML file sits in the tree. Moving a file without updating the fields changes nothing. Updating the fields without moving the file still works correctly. Always treat the fields below as the source of truth. The fields that actually determine placement: collection id (entity id of a collection) — places the entity in that collection. null or omitted → root collection. parent id on a collection — this, and only this, sets the collection's own parent. A collection's position in the folder tree is ignored on import; without parent id (or with parent id: null ) the collection becomes a root level collection, no matter how deep its folder is nested. To nest one collection under another, set parent id to the parent collection's entity id . dashboard id / document id on a card — nests a card under a dashboard or document. Such a card must also set collection id to match the parent's collection id . A card never sets both. On disk, cards nested under a dashboard or document live in a subfolder next to the parent YAML (e.g. my dashboard/card.yaml sitting next to my dashboard.yaml ) — but again, this is purely for human navigation; the fields are what Metabase reads. Import paths Metabase only imports YAML from these top level directories; anything outside is ignored: collections/ — all user content (cards, dashboards, documents, actions, snippets, transforms, etc.), partitioned by namespace: main/ , snippets/ , transforms/ , data actions/ , data apps/ . databases/ — the segments/ and measures/ subdirectories under each table, and each Library published table's metadata. python libraries/ (also accepted as python libraries/ ). glossary/ — glossary terms (not covered by the spec). transforms/ — contains transform jobs/ and transform tags/ . actions/ — legacy location for actions, still read on import. New exports write an action in its collection's directory, beside that collection's cards, not in an actions/ subdirectory. data apps/<slug / — a data app's data app.yaml and its built bundle. The app's collection, with its cards and actions, is under collections/data apps/ . serdes/meta Every entity carries a top level serdes/meta array with a single entry naming its model ; Metabase derives the identity from the entity's own fields ( entity id , or name and table id for a Field). Older exports carry the full identity path instead ( {id, model, label?} per entry), which Metabase still imports. Example: validate schema reads serdes/meta to determine which entity type each file represents. The full rules (including nested entities and composite identity paths) are in the spec. Reading the spec Extract the current spec from the package into a temporary file, and read it from there: Beyond the per entity shapes summarized in this SKILL, the spec also covers: MBQL query form (stages, field references, joins, expressions, aggregations, filter/expression operators, temporal bucketing, binning), native queries and template tags ( text , number , date , boolean , dimension , temporal unit , card , snippet , table ), visualization settings, click behavior, dashboard/card parameters, and the data app layout ( data apps/<slug /data app.yaml and the app's collection under collections/data apps/ ). Reach for it whenever edits touch any of those. Read on demand, not eagerly. Extract and open the spec only when you are about to read or modify content files for the entities listed above — e.g. the user asks to edit a card, add a dashcard, tweak a transform, or similar work that implies YAML edits. Do not extract it at session start or for tasks unrelated to representation YAML. Extract it once per session; it is long, so search it for the section you need rather than reading it whole. Validating Validate edits with the built in CLI: Pass the top level export folder, or the git repository root. The tool walks the import paths listed above, reads serdes/meta on each file to pick the right validation rules, and exits non zero on failure. Prefer running this over manually cross checking field shapes. It's essentially instant, so invoke it whenever useful — after each edit, between edits, whenever the shape of a file feels uncertain. No reason to batch. Generating entity IDs Every entity needs a 21 character NanoID for entity id . Generate one (or several) with the bundled CLI: Generating UUIDs Some fields in the format require v4 UUIDs rather than NanoIDs — notably lib/uuid on MBQL aggregation clauses (referenced from order by and later stages) and the id on dashboard/card parameters. Generate them with: