# RowCove product and capability map

## Who it is for

RowCove is for people who work with data large enough that waiting, copying or loading the whole file into a general-purpose tool wastes time:

- AI dataset engineers preparing instruction, preference, chat, evaluation and labeling records
- pre-training and post-training teams working with shards, splits and partitioned datasets
- quantitative researchers inspecting candles, trades, quotes, order books, funding, open interest and backtest outputs
- engineers reviewing multi-gigabyte CSV, JSONL, Parquet or Arrow exports
- teams that need a local, read-only-first view before they run a transformation elsewhere

## What makes the workflow different

### Automatic opening

One Open data entry point accepts a file or folder. The detector uses content evidence and container metadata, preserves suffix and compression information, guesses headers and delimiter from a bounded sample, and reports the detected reader before work starts. Dataset folders can expose splits, shards and Hive-style partition keys.

### Bounded browsing

The first page is intentionally bounded. Rows and cells are rendered through a virtualized data view, columns can be projected, long values remain inspectable and the application does not pretend that a preview is a full count. Large work shows progress, supports cancellation and reports recoverable errors.

### Full-file work

Supported sources can be filtered, sorted, profiled or checked without asking the UI to render the entire file. Large CSV query work uses a streamed source-row index and publishes results atomically; source identity and stable ordering are retained. Query pages carry source provenance so an agent can explain where a result came from.

### Reversible editing

The original source remains available. Supported edits are staged, recoverable and written with Save copy to a new path. A failed or cancelled write leaves the source unchanged. Format capabilities are shown before an operation so a read-only adapter is not mistaken for a writer.

### Local MCP

The desktop application and its local stdio MCP process share the same format detection, authorization and bounded data contracts. MCP is off until the user enables it, scopes allowed roots and gives the client the generated configuration. Tool schemas, test calls and call history are visible in the product.

## Honest boundaries

Capabilities vary by format, layout, device resources and plan. JSON is a bounded document reader; JSONL is streamed and indexed; Parquet, Arrow and Feather use column projection; spreadsheets are bounded and read-only for workbook writes. A complete scan, exact count, global sort or quality report must be requested and labeled as such. There is no universal speed guarantee and no claim that every TB-scale operation fits in memory.

## Formats at a glance

See the [full format matrix](./formats.md). In brief, RowCove recognizes delimited text, JSON/JSONL/NDJSON, Parquet, Arrow IPC/Feather, spreadsheets, compressed text, archives and WebDataset TAR, arrays and tensors, TFRecord, Safetensors, Zarr, SQLite, DuckDB, HDF5, Avro, ORC, XML, ODS, XLSB and MessagePack. Operation availability is reported per detected source.