Text NLP AI

Description

The Text NLP AI step runs one of seven per-row tasks over your data, each as a batch of structured (forced tool-use) LLM calls: Named Entity Recognition, POS Tagging, Summary, Zero-shot Classification, Key/Value Extraction, Prompt, and Embed. It runs as a managed PlaidCloud job (like ML: Train Model and the LLM Step), so it does not compete with the rest of the workflow for resources.

Because every task is a real LLM call, results are not bit-for-bit reproducible between runs — see Non-Determinism.

Configuration

• Source / Output tables — the input table and the table the results are written to.
• Text Column (field) — the column each task reads.
• Task (task) — one of ner, pos, summary, zeroshot, keyvalue, prompt. The embed task is not in the form’s task list; set it with an AI assistant or the REST API.
• Language (language) — the text’s language, passed to the model. Defaults to en.
• LLM Connection (llm_connection_id) — an LLM connection for Anthropic, OpenAI, Gemini, Azure OpenAI, Grok, or GitHub Models. Leave blank to use the workspace default.
• Model (llm_model) — leave blank to use the connection’s default model.
• Temperature (temperature) — optional sampling temperature. See Non-Determinism: current-generation models ignore it.
• Max rows (max_rows) — answer only the first N rows in source order. The rest keep an empty answer. Leave blank for every row.
• Batch Size (batch_size) — rows sent per LLM call, 1 to 500. Defaults to 25. A batch whose response is truncated is split in half and retried rather than losing rows.

Tasks

Four tasks return one output row per input row (row-preserving); three return zero or more rows per input row and carry the source’s identifier columns (id_columns) onto each one (row-exploding).

Task	Fan-out	Output columns	Task-specific fields
Named Entity Recognition (ner)	Row-exploding	entity_text, entity_type	Entity Types (entity_types) — the allowed type list. Defaults to people, organizations, locations, dates, quantities if left blank.
POS Tagger (pos)	Row-exploding	token, pos (universal POS tag, for example NOUN, VERB, ADJ)	None.
Summary (summary)	Row-preserving	summary	Summary Length (summary_length) — maximum sentences, 1 to 20. Defaults to 3.
Zero-shot Classification (zeroshot)	Row-preserving	label, score (0 to 1)	Label Field (label_field) — the source column holding each row’s candidate labels; without it the model has nothing to choose from. Multi-Label (multi) — allow more than one label per row (joined with , ) instead of picking the single best.
Prompt (prompt)	Row-preserving	answer, error	Prompt (prompt) — the instruction applied to every row, required. Input Fields (input_fields) — the columns sent to the model with the prompt, as one record per row; blank sends the Text Column.
Embed (embed)	Row-preserving	embedding (a JSON array of numbers), error	None. Set with an AI assistant or the REST API. Needs an OpenAI-compatible or Azure OpenAI connection, and the step’s Model must name an embedding model, since the connection’s default is usually a chat model; an Anthropic connection is refused because it has no embeddings API. A failed request keeps every row of its batch, with an empty embedding and the reason in error.
Key/Value Pairs (keyvalue)	Row-exploding	key, value	Key Field (key_field) — the source column holding each row’s candidate keys to extract; leave blank to let the model find keys freely. Fuzzy (fuzzy) — match candidate keys loosely instead of exactly.

ID Columns (id_columns) carry identifying columns from the source table onto every output row of a row-exploding task, so you can trace an extracted entity or pair back to its source record. An ID column can’t share its name with one of the task’s own output columns: naming an ID column entity_type on an ner task, for example, is refused when you save the step, with ID column ‘entity_type’ has the name of a column the ner task writes; rename or remove it.

The error column holds the message of the call that failed for that row. The Prompt task is for any per-row question the other tasks don’t cover, such as Classify the sentiment of this review. Field values are sent as data, never as instructions to the model.

For the Prompt task, Document column (content_field) names a column that holds each row’s document as base64. The step then sends one document per call. The document goes to the provider as an image or PDF when the provider accepts that type and size. Otherwise the step sends the text in Text fallback column (text_fallback_field), such as OCR text. Whole documents leave PlaidCloud for the provider, so check that this is acceptable for your data. Rows are sent one at a time, so Batch Size is ignored. If a batch fails, its rows are kept with an empty answer and the step warns.

Saving is also refused without a source table (Source table is required.), or with a Batch Size or Summary Length outside the ranges above.

Non-Determinism (LLM-Backed Results)

Every task is a real call to an LLM, so results can vary between runs even with identical input — there is no bit-for-bit guarantee, and this is expected, not a bug.

Temperature exposes a further wrinkle: current-generation models (Opus 4.7+, Sonnet 5+, Fable) reject an explicit sampling temperature outright. If you set one, the step tries it once; the first rejection is detected, the run falls back to the model’s own server default for the rest of the batches, and the step’s result carries a one-time informational note rather than failing. Leaving Temperature blank (the default) skips this round-trip entirely and is the recommended setting on current models.

Related

• ML: Train Model — fit a model rather than call an LLM per row.
• LLM Step — a free-form agentic LLM call with scoped read/write project access, rather than one of these fixed per-row tasks.
• Alteryx Conversion Matrix — Named Entity Recognition, POS Tagger, Text Summary, Zero-shot Text Classification, and Key/Value Pairs all convert to this step.
• Migrate Savant Analyses — Savant Infer and Vision agents convert to the Prompt task.