Datasets and Frames

A dataset is the input to a pipeline. Each item in it is one row: an object in a JSON list, a row in a CSV or Parquet file, or one file in a directory.

In the Python API, a Frame is the data structure that represents a dataset. Readers return a Frame, and every operation on a Frame returns a new Frame representing the transformed dataset. See Frame methods below.

Defining a dataset

YAMLPython

Datasets are declared at the top level of the config and referenced by name in the pipeline's steps:

datasets:
  user_logs:
    type: file
    path: "user_logs.json"

A reader loads a dataset and returns a Frame:

import docetl

user_logs = docetl.read_json("user_logs.json")   # or read_csv, read_parquet, read_dir
in_memory = docetl.from_list([{"text": "..."}])  # from a list of dicts

Chaining operations like .map() records them without running anything; execution happens at an action like .collect(). See Frame methods.

Examples

A JSON file

A list of objects; each object is one row.

// reviews.json
[
  {"id": 1, "product": "headphones", "review": "Battery died after a week."},
  {"id": 2, "product": "keyboard", "review": "Keys feel great, very quiet."}
]

YAMLPython

datasets:
  reviews:
    type: file
    path: "reviews.json"

reviews = docetl.read_json("reviews.json")

A CSV or Parquet file

Each row of the table is one row of the dataset; column names become keys.

ticket_id,customer,message
101,acme,"Cannot log in since the update"
102,globex,"Invoice total looks wrong"

YAMLPython

datasets:
  tickets:
    type: file
    path: "tickets.csv"   # or .parquet

tickets = docetl.read_csv("tickets.csv")   # or read_parquet(...)

A directory of documents

Every non-hidden file under the directory (recursively) becomes one row: text holds the file's content, with filename and path alongside. PDF, Word, PowerPoint, and Excel files are converted to text; other files are read as UTF-8; binary files with no extractor are skipped with a warning.

contracts/
  acme_msa.pdf
  globex_nda.docx
  notes/renewal_2026.txt

YAMLPython

datasets:
  contracts:
    type: file
    path: "contracts"

contracts = docetl.read_dir("contracts")
# one row per file:
# {
#     "filename": "acme_msa.pdf",
#     "path": "contracts/acme_msa.pdf",
#     "text": "MASTER SERVICE AGREEMENT\nThis Agreement is entered into by...",
# }

An in-memory list (Python only)

docs = docetl.from_list([
    {"speaker": "patient", "utterance": "The headaches started last month."},
    {"speaker": "doctor", "utterance": "Any changes in vision?"},
])

Relative paths resolve against the directory you run from, not the location of the YAML file or Python script.

Frame methods

A Frame is lazy and immutable: each operation records a step and returns a new Frame, and nothing runs until an action.

Readers create a Frame: docetl.read_json, read_csv, read_parquet, read_dir, from_list, and Frame.from_yaml.
Operations return a new Frame: map, filter, reduce, resolve, equijoin, extract, split, gather, unnest, cluster, sample, parallel_map, and the code variants code_map, code_filter, and code_reduce.
Actions run the pipeline: collect() (rows as a list of dicts), to_pandas() (a pandas DataFrame), show() (run on a few rows and print), count(), and write_json() / write_csv() / write_parquet().
Inspection and export: schema() (output fields, computed without running), total_cost and token_usage (after a run), to_yaml() and to_python() (export the pipeline), and optimize() (run the MOAR optimizer).

See the Python API reference for full signatures.

Parsing tools (non-standard inputs)

DocETL ships built-in parsing functions for file types beyond the above, e.g., whisper_speech_to_text for audio, and you can register your own. See Custom Parsing for the available built-in tools and how to define custom ones.

To use one, point the dataset at a JSON file of paths and attach the parsing function:

YAMLPython

datasets:
  audio_transcripts:
    type: file
    source: local
    path: "audio_files/audio_paths.json"
    parsing_tools:
      - input_key: audio_path
        function: whisper_speech_to_text
        output_key: transcript

import docetl

audio_transcripts = docetl.read_json(
    "audio_files/audio_paths.json",
    parsing=[
        {
            "input_key": "audio_path",
            "function": "whisper_speech_to_text",
            "output_key": "transcript",
        }
    ],
)

input_key: the key holding the path to the file to parse.
function: the parsing function (built-in or custom).
output_key: the key the parsed content is stored under, accessible in prompts as {{ input.transcript }}.