Configuration Reference

Where settings live

Every way of starting a run (a file upload on the Pipeline tab, a connector run, a folder or cloud listener, an SFTP inbox, a scheduled run, a trigger) opens the same Configure window, so a setting means the same thing and has the same default everywhere. Its controls are grouped into three tiers: Common (most runs touch them), Advanced (for datasets that misbehave) and Experimental (rarely right; all off by default).

Transport	Spelling	Used by
JSON	snake_case, for example text_mode	Stored pipeline configs of schedules, listeners and SFTP inboxes, trigger target_pipeline_config, POST /api/v1/process-connector. camelCase is also accepted; snake_case wins.
Multipart form	camelCase, for example textMode	The Pipeline tab upload, POST /process-data. Lists and objects are sent as JSON strings.
config field	its own layout	POST /api/v1/process and the Python SDK; see the last section.

Settings marked (Enterprise) are shown only to Enterprise accounts and are ignored by the server for every other account, without an error. Speed-preset knobs, Enterprise internals and module logic are outside the scope of this page; see Pipeline modules for what each stage does with these settings.

Targets, columns and size

Setting (control)	JSON key	Multipart field	Default	Options / range	What it does
Target Columns (Y Variables)	y_columns	yColumns	none (required)	column names	The columns to predict; several targets of mixed task types are allowed. Replay runs take them from the parent session.
Excluded Columns	excluded_columns	excludedColumns	none	column names	Columns that are not features. They stay in the outputs untouched: not encoded, imputed, scaled or removed.
Task Type Selector	task_types	taskTypes	regression on the Pipeline tab, classification on automation	c or r per target, in target order	Classification or regression. /api/v1/process treats every target as classification.
Target Rows / Output Sample Size	output_number (also output_rows)	outputNumber	10000	positive whole number, or empty	Size of the final dataset. Empty skips synthesis. Zero, negative or unreadable values fall back to 10000.
Speed / accuracy	speed_mode	mode	balanced	rapid, balanced, accurate	Speed preset; see Speed modes.

In JSON a bare mode is the run mode (full_pipeline, inference, retraining), never the speed preset.

Data cleaning (Anomaly Detection)

The Data cleaning section lists fourteen operations; the stage runs when at least one is ticked. In JSON the stage switch is enable_anomaly_detection (default false; multipart enableAnomalyDetection) and the operations are the anomaly_params object (multipart anomalyParams). Missing keys take the defaults below; a request with no anomaly_params uses the operations saved on your account.

Control	anomaly_params key	Default	Effect
Currency conversion	currency_conversion	on	Currency amounts become USD numbers.
Percentage to decimal	numeric_range_conversion	on	"85%" becomes 0.85.
Date/time standardisation	date_time_conversion	on	One date and time format per column.
Whitespace cleaning	whitespace_cleaning	on	Extra spaces removed.
Boolean normalisation	boolean_normalization	on	Yes/No, 1/0, true/false unified.
Written numbers to digits	textual_numeric_conversion	on	"five" becomes 5.
Ordinal detection	ordinal_detection	on	Ordered categories such as low, medium, high are recognised.
Spelling error detection	spelling_error_detection	on	Misspelt category values corrected.
Unit standardisation	unit_standardization	off	"5kg" becomes 5 with the unit recorded.
Data type consistency	data_type_consistency	off	One type per column.
Phone area code	phone_area_code_extraction	off	Adds an area-code column.
Email domain	email_domain_extraction	off	Adds an e-mail domain column.
Emoji and symbols	special_character_analysis	off	Emoji converted to text.
(API and automation only)	temporal_feature_extraction	off	Time features derived from dates.

Instructions, completion and run options

Setting (control)	JSON key	Multipart field	Default	Options / range	What it does
Instructions for the pipeline	prompt_text	promptText (also projectDescription)	empty	text	Plain-language context for the cleaning stages. Experimental.
Save artifacts for inference / retraining	retraining	retraining	false	true / false	A label on the run only: every run keeps its replay artifacts, and any run whose artifacts are still stored can be a parent.
Session name	session_name	sessionName	none	text	Name shown in Runs. No Pipeline tab field; rename from Runs.
(no control)	text_cleaning	textCleaning	true	true / false	Reserved; not available yet.
(no control)	llm_enabled	not read	true	true / false	false: no stage contacts a language model for this run; completion and validation are skipped. JSON surfaces and advanced_params on /api/v1/process.
(no control)	strict_llm	not read	false	true / false	true: a stage that could not reach its model fails the run instead of falling back.

Dataset completion & validation

JSON completion_params, multipart completionParams. Not read by /api/v1/process.

Field	Default	Options	Effect
completion_enabled, validation_enabled	false, false	true / false	Fill empty cells; check existing values. The stage runs when either is on.
completion_mode, validation_mode	fast, fast	fast, accurate	Accurate cross-checks more sources and is slower.
validation_action	overwrite from the dashboard, mark when omitted	overwrite, mark	Replace wrong values, or flag them in a new column.
filling_columns, validation_columns	all	column names	Limit each half to these columns.
confidence_threshold	0.3	0 – 1	Values found with lower confidence are not written.
api_usage_per_cell	0.01	number	Reserved; not available yet.

Stages and outputs (Output & Tool Selector)

One dialog chooses which stages run, which files they write, the export format, file names, the model-ready export and the PDF report. It travels as output_preferences (multipart outputPreferences). Automation payloads also accept enable_dtc, enable_mdh, enable_cds, enable_dsm, enable_dsg (default true) and enable_anomaly_detection (default false); a key in output_preferences wins. Unsent keys use your saved account defaults.

Selector option	output_preferences key(s)	Default	Effect
Data Completion & Verification	completion_use_tool, compleated_dataset_output	off	Run DCV; keep compleated_dataset.csv.
Anomaly Detection	anormaly_detection_use_tool, anormaly_fixed_output	off	Derived from Data cleaning; keep anormaly_fixed_output.
DTC, MDH, CDS, DSM, DSG	<stage>_use_tool, <stage>_output	on, on	Run the stage; keep its file.
(hidden) DOR	dor_use_tool, dor_output	on	Applies to retraining only; outlier removal does not run in training runs.
Write every output as	output_format	csv	csv, parquet, feather, orc, xlsx, json, jsonl, ndjson	Format of every data output. An output Excel cannot hold, or a failed write, falls back to CSV and the run says so. Inference/retraining results, compleated_dataset and feature_selection_output are always CSV.
File names	output_name_pattern	prefixed	prefixed, plain	prefixed: sales_20260922_013000_dsg_output.csv; plain: dsg_output.csv. Automated runs omit the input name.
Multimodal Output	multimodal_output	off in the dashboard	true / false	Write the merged text/media frame.
Pipeline Report PDF	pipeline_report_output	on	true / false	Write pipeline_report.pdf. Automation may send enable_pdf_report, which wins.
Model-Ready Dataset	model_ready_output	on	true / false	Write the all-numeric export.
Split layout	model_ready_split	none	none, xy, train_test, xy_train_test	Single file, X and y, train and test, or all four.
Test size	model_ready_test_size	0.2	0.05 – 0.5	Fraction held out.
Test set from real rows only	model_ready_real_test_only	on	true / false	Off splits all rows naively and adds a leakage warning.

Common settings

Setting (control)	JSON key	Multipart field	Default	Options / range	What it does
Feature Selection, before imputation	enable_feature_selection	enableFeatureSelection	off	true / false	Drops the least useful columns right after DTC; targets and excluded columns are kept. See Feature selection.
Method / Keep top N	feature_selection_config {method, top_k, threshold}	featureSelectionConfig	mutual_information, 20	mutual_information, shap_importance, rfe, variance_threshold, correlation_pruning, pca (API also permutation_importance, leakage_guard, a list, auto); top N 1 – 1000	How columns are scored and how many are kept. Variance and correlation use threshold (0.01 / 0.95) instead of top N. Omitting method in JSON means variance_threshold.
Text Tokenization	text_mode	textMode	2	0 Drop Text, 1 Neural Tokenization, 2 TF-IDF, 3 Auto	How free-text columns become numbers. Mode 0 does not remove text: those columns are carried through unencoded.
Output Sample Size	output_number	outputNumber	10000	as Target Rows	Same value as Target Rows.
Audio / Image column detection	(dashboard only)	(dashboard only)	auto	auto, manual	Detect media columns, or pick them yourself.
Select Audio / Image Columns	audio_columns, image_columns	audioColumns, imageColumns	none	column names	Columns holding media file names. Automation never detects media on its own.
Enable Preliminary Cleaning	enable_preliminary_cleaning	enablePreliminaryCleaning	on	true / false	Removes exact duplicate rows, turns tokens such as na and null into empty cells, drops nearly-empty, constant and ID-like columns. Target, excluded and media columns are protected.
Target model family (Enterprise)	model_families	modelFamilies	linear + tree_gbdt (one pre-selected on automation forms)	linear, tree_gbdt, mlp, sequence_neural; up to 4	Prepares the data for a model family; with several, prepares once per family and recommends one.

Advanced settings

Setting (control)	JSON key	Multipart field	Default	Options / range	What it does
MDH Mode	mdh_mode	mdhMode	0	0 Imputation, 2 Imputation/dropping (Enterprise), 1 2D-Removal	Imputation fills every gap and keeps every row; Imputation/dropping removes what is too empty to fill reliably and fills the rest; 2D-Removal removes rows and columns with too many gaps.
Audio Mode	audio_processing_mode	audioMode	0	0 No Processing, 1 Basic Features, 2 Advanced Processing, 3 Mel-Spectrogram Tokenization	Encoding of audio files. Out-of-range values mean 0.
Image Mode	image_processing_mode	imageMode	0	0 No Processing, 1 Basic Features, 2 Advanced Processing, 3 Neural Embeddings	Encoding of image files.
Scaling Method	cds_scaler	cdsScaler	auto	auto, standard, minmax, robust, maxabs, yeojohnson, quantile_normal, quantile_uniform, log1p, none	auto picks the best-scoring scaler; anything else forces one.
Data Similarity Threshold	similarity_p	similarityP	0.99	greater than 0, less than 1	DSM keeps one column of any pair more similar than this; lower removes more columns. DSM never splits rows. Values above 1 are read as percentages; invalid values fall back to 0.99.
Z-Score Limit	zscore_limit	zscoreLimit	3	greater than 0	Extreme values are capped at this many standard deviations while the generator learns; lower gives fewer extreme synthetic values.
Synthesis Mode	dsg_mode	dsgMode	copula	copula, gan	GAN captures complex relationships but is much slower. A Pipeline tab upload sends only an explicit GAN, so leaving Copula there lets AutoData choose.
Class Balancing: Minimum Class Share	dsg_y_oversample_threshold	dsgYOversampleThreshold	empty (off)	greater than 0, less than 1	Raises every target class (or bin of a continuous target) to at least this share with synthetic rows; the output can exceed Target Rows. Out of range returns 400.
Class Balancing: bins	dsg_y_num_bins	dsgYNumBins	10	2 – 100	Equal-frequency bins for a continuous target.
Allow Stratified Downsampling	dsg_allow_downsample	dsgAllowDownsample	off	true / false	When Target Rows is below the processed rows: a class-preserving subsample (on) or the full processed data without synthesis (off).
Pivot (as-of) column (Enterprise)	pivot_config	pivotConfig	off	column; leakage_action warn / relative / drop; relative_deltas; auto_suggest	When each row came into existence; dates are measured from it and later-recorded columns are handled as chosen.
Validation method (Enterprise)	validation_policy	validationPolicy	off	split_strategy random / stratified / time_ordered / grouped; eval_protocol holdout / kfold / walk_forward; order_column, group_column, eval_folds (3), walk_forward_splits (3)	How rows are divided when preparation choices are scored.
Dataset metadata (Enterprise)	dataset_metadata	datasetMetadata	none	JSON dictionary (version 1) or CSV data dictionary	Declared roles, types, ranges, order and missing-value sentinels override detection in every stage and are replayed at inference. Unreadable metadata returns 400.

Experimental settings

Setting (control)	JSON key	Multipart field	Default	Options / range	What it does
Force Pure Synthetic Output	force_synthetic	forceSynthetic	off	true / false	All rows synthetic. Honoured only on Pipeline tab uploads; other surfaces accept and ignore it.
Result Evaluator	enable_result_evaluator	enableResultEvaluator	off	true / false	Benchmarks several models on the prepared data; writes result_evaluation_report.json.
Column selection, after scaling (Enterprise)	selection_config	selectionConfig	off	methods (leakage_guard, variance_threshold, correlation_pruning, mutual_information, permutation_importance, rfe, shap_importance), top_k, threshold, leakage_action warn / drop	Target-aware selection inside DSM.
Probe regularization (Enterprise)	probe_config	probeConfig	l2	penalty l2 / l1 / elasticnet / none; C; l1_ratio	Regularisation of the model that scores preparation choices, not of your model.
Column decisions (Enterprise)	imputation_overrides, scaling_overrides	imputationOverrides, scalingOverrides	none	per column, e.g. {"age": "Median Imputation"}	Pin a column's imputation or scaling; a pin that cannot be honoured is refused and recorded.

Settings without a control

Setting (control)	JSON key	Multipart field	Default	Options / range	What it does
Datetime expansion	datetime_mode	datetimeMode	basic (or the speed mode's value)	none, minimal, basic, full	Columns per date: none drops dates, minimal keeps elapsed days, basic adds month, day of week, year and hour, full adds every component. Not read by /api/v1/process.
Nested flattening	enable_nested_flattening	enableNestedFlattening	on	true / false	Expand JSON objects in cells into columns; lists stay as text.
Columns to flatten	flatten_columns	flattenColumns	absent (detect)	list, or []	Only these columns; [] expands none.
Rare-class oversampling	dsg_rare_oversample_threshold	dsgRareOversampleThreshold	none	0 – 0.5	Reserved; not available yet. Use Class Balancing.
Idempotency key	header Idempotency-Key	idempotencyKey	none	string	/process-data only: a repeated key returns the existing session.
Webhooks	webhook_ids	webhookIds	none	webhook ids	Webhooks notified when the run finishes or fails.
Automatic write-back	auto_sink	not read	none	object	Automation and /api/v1/process: where results are written on completion.
Media files	n/a	imageFiles, soundFiles	none	files	/process-data only: the files named by the media columns.
Schema / ranges	n/a	schema, ranges	none	JSON	Reserved; not available yet.

Speed modes

rapid finishes sooner with rougher choices, balanced (default) is the standard behaviour, accurate takes longer to fit your data more closely. The price does not change with the mode, only time. Modes do not apply to inference or retraining. Each mode sets sixteen knobs, all overridable:

Knob	Override key	Range	Rapid / Balanced / Accurate	What it changes
One-hot / free-text cardinality limit	dtc_OHElimit	2 – 100000	40 / 100 / 250	Which columns count as free text rather than categories.
Text-detection sample rows	dtc_sample_rows_detect_text	10 – 100000	100 / 200 / 400	Rows read when detecting text columns.
Auto text-mode sample rows	dtc_sample_rows_text_mode	10 – 100000	100 / 300 / 600	Rows read when Auto picks TF-IDF or neural.
Neural tokenizer vocabulary size	dtc_seq_max_features	100 – 1000000	20000 / 50000 / 50000	Vocabulary cap, text mode 1.
TF-IDF vocabulary size	dtc_tfidf_max_features	100 – 1000000	8000 / 20000 / 20000	TF-IDF feature cap, text mode 2.
Datetime expansion	dtc_datetime_mode	none, minimal, basic, full	minimal / basic / full	Columns per date.
Synthetic text method	text_synth_method	statistical, vae	statistical / statistical / vae	Generator for synthetic text.
Anomaly detection: skip LLM checks	anomaly_force_no_llm	true / false	true / false / false	LLM-backed anomaly checks off.
Anomaly detection: sample scale	anomaly_sample_scale	0.1 – 10	0.5 / 1.0 / 2.0	Rows each anomaly check inspects.
Missing-data threshold trials	mdh_nooftrials	1 – 100	2 / 5 / 10	Thresholds tried by 2D-Removal.
Skip expensive imputation strategies	imputer_skip_expensive	true / false	true / false / false	Slowest imputation methods left out.
Force all imputation strategies	imputer_force_all	true / false	false / false / true	Every imputation method tried.
Scalers evaluated	cds_eval_top_n	1 – 9	1 / 3 / 9	Candidate scalers tried under auto.
Synthetic-generation fit budget	dsg_iess_budget_multiplier	0.1 – 10	0.5 / 1.0 / 2.0	How much data the generator fits on.
Synthetic image method	image_synth_method	statistical, vae	statistical / statistical / vae	Generator for synthetic image features.
Synthetic audio method	audio_synth_method	statistical, vae	statistical / statistical / vae	Generator for synthetic audio features.

Out-of-range values are clamped; unreadable ones are ignored.

Key (JSON / multipart)	Meaning
mode_overrides / modeOverrides	Your values for the knobs, keyed {"dtc.OHElimit": 250} or {"dtc_OHElimit": 250}. On /api/v1/process put the flat keys directly in advanced_params.
mode_decisions / modeDecisions	Per setting, user (keep yours, or the Balanced value if you set none) or mode (preset wins). Keys are groups text, datetime, missing_data, scaling, anomaly, synthetic, media, or single knobs such as dtc.OHElimit.
touched_fields / touchedFields	Which settings you changed yourself; carried with the run.

Without a decision, a value you set wins; otherwise the preset. The Pipeline tab asks in its Confirm processing run dialog (Keep my settings / Use <mode>); automation forms ask once in the Configure window and store the answer; /api/v1/process returns 409 mode_confirmation_required with mode_contested and policy_signature. Re-send with advanced_params.mode_ack set to the signature to keep your values, or with advanced_params.mode_decisions. The SDK raises ModeConfirmationRequired.

Names on POST /api/v1/process

This route takes a multipart file and a config JSON field:

{
  "target_columns": ["churn"],
  "output_rows": 10000,
  "session_name": "churn-weekly",
  "tools": {"anomaly": false, "dtc": true, "mdh": true, "dor": false,
            "cds": true, "dsm": true, "dsg": true,
            "feature_selection": false, "preliminary_cleaning": true,
            "nested_flattening": true, "result_evaluator": false},
  "feature_selection_config": {"method": "mutual_information", "top_k": 20},
  "anomaly_params": {"currency_conversion": true},
  "advanced_params": {
    "excluded_columns": ["id"], "text_mode": 2, "mdh_mode": 0,
    "cds_scaler": "auto", "similarity_p": 0.99, "zscore_limit": 3,
    "dsg_mode": "copula", "dsg_y_oversample_threshold": 0.3,
    "dsg_y_num_bins": 10, "dsg_allow_downsample": false,
    "audio_mode": 0, "audio_columns": [], "image_mode": 0, "image_columns": [],
    "output_format": "parquet", "output_name_pattern": "prefixed",
    "llm_enabled": true, "strict_llm": false,
    "mode": "balanced", "mode_decisions": {}, "touched_fields": []
  },
  "webhook_ids": [], "auto_sink": null
}

On this route advanced_params.mode is the speed mode, media modes are audio_mode / image_mode, every target is classification, the model-ready split comes from your saved defaults, run_dsg: false skips synthesis, and completion_params, datetime_mode and force_synthetic are not read. Enterprise keys go in advanced_params.