File I/O

Read and write local files — CSV, JSON, NDJSON, Parquet.

Supported formats

Format	Extensions	Read	Write	Notes
CSV	.csv, .tsv	Yes	Yes	Configurable delimiter, headers, encoding.
JSON	.json	Yes	Yes	Array of objects or single object.
NDJSON	.ndjson, .jsonl	Yes	Yes	One JSON object per line. Native interchange format.
Parquet	.parquet	Yes	No	Columnar binary format. Read via DuckDB.

Config fields

Field	Required	Default	Description
path	Yes	—	File path, relative to project root.
format	No	Inferred from extension	One of csv, json, ndjson, parquet.
encoding	No	utf-8	File encoding. Applies to CSV and JSON.
headers	No	true	Whether the first row contains column names. CSV only.
options.delimiter	No	,	Column separator for CSV/TSV.

Paths are resolved relative to the project root (the directory containing flow.yaml). Absolute paths are rejected.

Reading files

CSV with headers

flow.yaml

read-leads:
  type: source
  op: file.read
  params:
    path: data/leads.csv
    format: csv
    options:
      delimiter: ","
      header: true
      encoding: utf-8
  outputs:
    rows:
      type: Table
      schema:
        name: { type: string }
        email: { type: string }
        score: { type: number }

JSON array

read-config:
  type: source
  op: file.read
  params:
    path: config/settings.json
    format: json
  outputs:
    settings: { type: Record }

NDJSON

read-events:
  type: source
  op: file.read
  params:
    path: logs/events.ndjson
  outputs:
    events: { type: Table }

Parquet

read-warehouse:
  type: source
  op: file.read
  params:
    path: warehouse/orders.parquet
  outputs:
    orders: { type: Table }

Writing files

Output to CSV

write-results:
  type: deterministic
  op: file.write
  params:
    path: output/qualified.csv
    format: csv
    options: { delimiter: ",", header: true }
  inputs:
    data: { type: Table, from: ref(filter-leads.qualified) }

Output to JSON

write-json:
  type: deterministic
  op: file.write
  params:
    path: output/report.json
    format: json
  inputs:
    data: { type: Table, from: ref(score-leads.scored) }

You can also write NDJSON by setting format: ndjson. The same options fields apply across all formats.

File path resolution

All paths are relative to the project root — the directory that contains your flow.yaml. Given this project structure:

my-project/
  flow.yaml
  data/
    leads.csv
  output/

A path of data/leads.csv resolves to my-project/data/leads.csv. Radhflow creates output directories automatically if they don’t exist.

Tip

NDJSON is the most efficient format for large datasets. Each line is an independent JSON object, so Radhflow can stream rows without loading the entire file into memory. If you’re working with more than 100k rows, prefer NDJSON over JSON.

Troubleshooting

File not found. Check that the path is relative to the project root, not your current working directory. Run rf validate to catch path issues before execution.

Encoding issues. If you see garbled characters, set encoding explicitly. Common values: utf-8, utf-16, latin-1, ascii.

Permission errors. Radhflow runs in a sandbox. The file must be inside the project directory or a path explicitly mounted into the runtime. Files outside the project root are inaccessible by design.

CSV delimiter mismatch. If your CSV uses tabs, semicolons, or pipes as separators, set options.delimiter to the correct character. TSV files (.tsv) default to tab.