# RowCove format and operation matrix

This matrix describes the public capability contract. “Inspect” means bounded metadata or a first page. “Query” means the reader exposes supported filters, projection, pagination or sorting; it does not promise that every operation is available for every file layout.

| Family | Formats | Current workflow |
| --- | --- | --- |
| Delimited | CSV, TSV, DSV | Detect delimiter and headers, index, page, project, filter, sort, edit and Save copy |
| Text | TXT, LOG | Bounded line records, sparse line index, page, filter, sort and new text artifact |
| Structured | JSON | Bounded document preview, object/array inspection and projection; read-only |
| Streaming records | JSONL, NDJSON | Stream index, page, dotted nested paths, projection, filter, bounded sort, edit and Save copy |
| Columnar | Parquet | Schema, column projection, partition-aware dataset browsing, bounded filter/sort and new artifact |
| Columnar | Arrow IPC, Feather | Schema, projection, bounded filter/sort and new artifact |
| Spreadsheets | XLSX, XLS | Workbook and sheet discovery, typed bounded pages and projection; workbook writes are read-only |
| Compression | GZIP, Zstandard, BZIP2 | Bounded streaming preview; complete materialization to a new destination with progress and cancellation |
| Archives | ZIP | Safe member catalog and supported entry browsing |
| Dataset archives | WebDataset TAR and compressed TAR | Sample member catalog and sequential sample browsing |
| Arrays | NPY, NPZ | dtype, shape, order, member metadata and bounded slices |
| Tensor/data formats | Zarr, Safetensors | Array or tensor metadata, ranges, chunks or bounded views |
| Training records | TFRecord | Sequential records, payload length, CRC status and bounded byte preview |
| Databases | SQLite, DuckDB | Explicit schema/table choice, projected read-only pages, parameterized filters and ordering |
| Metadata or specialized | HDF5, Avro, ORC, XML, ODS, XLSB, MessagePack | Bounded metadata, text or member views; native typed editing varies by format |

## Dataset-folder behavior

JSONL, Parquet and Arrow folders are treated as datasets. Discovery can expose split/config/shard names, file formats and Hive-style partition keys such as 'symbol=BTCUSDT/year=2026/month=10'. Use exact partition values and half-open ranges when the task depends on time or symbol. Read only the columns needed for the question.

## AI and quantitative shapes

Common AI fields include 'prompt', 'completion', 'chosen', 'rejected', 'messages', 'label', 'input_ids', masks and metadata. Common quantitative roles include timestamp, symbol, venue, interval, open, high, low, close, volume, trade count, side, price, size, funding rate and open interest. Preserve raw timestamps, decimal identifiers, units, timezone, split boundaries and source-row IDs. Do not silently coerce fixed-point prices or large identifiers to imprecise floating point.

## Format questions an agent should ask

1. Is this a single file, a directory of shards or a database?
2. Which split, symbol, date range or partition is in scope?
3. Which fields are required in the answer?
4. Is the requested result a preview, a complete scan, a transformation or a write?
5. Does the user need a new artifact, and where may it be created?

The detected source report is authoritative for the current file. Do not route an unknown or corrupt container through a text parser merely because its suffix looks familiar.