Studio · Full-screen workspace

Datasets

The Datasets studio turns a spreadsheet into something you can query like a database. Upload a CSV or Excel file and it becomes a set of tables with an inferred schema; from there you filter, sort, group and aggregate — all computed locally, with zero model tokens, so the numbers are exact. Re-upload a newer version and it tells you exactly which rows changed.

Your spreadsheet, but structured. Instead of pasting a table into chat and hoping the model reads it right, a dataset is parsed once: column types are inferred, low-cardinality columns become facets you can filter on, and a natural row key is detected. Every query after that runs on the real data — no estimation, no token cost.

▦ Tables from files

CSV, TSV, TXT and Excel — each sheet becomes a queryable table.

▦ Query, zero tokens

Filter, sort, group and aggregate — computed locally, exactly, for free.

▦ Real schema

Inferred types, low-cardinality facets with values, and a detected row key.

▦ Version diffs

Re-upload and see which rows were added, removed or changed.

Open it any time — say "open Datasets" or click the Open chip — or jump straight to a dataset. Cody works your datasets from chat, so you can ask questions of a table without opening the studio at all.

What it is#

A dataset is one uploaded file that holds one or more tables (a spreadsheet with several sheets becomes several tables). The studio stores it, parses its structure, and gives you a database-like surface over it: inspect the schema, run queries, and compare versions. It's the right home for the exports and tracking sheets that would otherwise live as loose files — a cost export, a headcount plan, a survey dump — anything you want to ask questions of rather than just read.

Getting data in#

Bring a file in and it's parsed into tables — CSV, TSV, TXT and XLSX are all supported, with dates normalised so downstream queries behave consistently. When a table already exists in the conversation — one you pulled from a doc or the wiki — an agent can hand you a one-click Save to Datasets card that files it as a dataset without you leaving chat.

Try saying

open Datasets save this table as a dataset import this cost export into datasets

The schema#

Every dataset carries an inferred schema so queries use real column names and valid values. For each table it records the columns and their types, flags low-cardinality columns as facets and lists their values (so you know a Team column has exactly "Platform / Infra / Data" to filter by), and detects a natural row key that identifies each row. That's what makes the diff meaningful and the queries precise.

Try saying

what columns and values are in this dataset? what can I filter on in the headcount table?

Querying like a database#

Query one table the way you would a database: filter on columns, free-text match across it, sort, and group-by / aggregate to get counts and totals. It returns matching rows, or grouped aggregates, and — crucially — it's computed locally with zero model tokens, so a "sum spend by team" or "count open rows per owner" is exact and costs nothing. Ask it a hundred ways and it stays free.

Try saying

group this cost export by team and sum the spend show rows where status is open, sorted by due date count tickets per owner

Versions & diffs#

Re-upload a file with the same name (or into the same dataset) and it appends a new version rather than overwriting — and computes a keyed row diff against the previous one. So you get, per table, exactly which rows were added, removed, or changed (with the fields that changed), plus any columns added or dropped. It defaults to the latest version versus the one before it, so "what changed since last week's upload" is one question.

Try saying

what changed between this week's and last week's upload? which rows were added in the new version?

Working from chat#

You rarely have to open the studio to use a dataset. Cody (and Sage) query your datasets directly — inspecting the schema, running a grouped query, or diffing versions — and answer in chat, grounded in the real rows. The studio is there when you want to browse the tables yourself; the data is queryable either way.

Try saying

what's the total spend in the AWS export? diff the latest headcount dataset vs the prior one