Troubleshooting text

Most surprises with text workflows fall into a few families: a model’s answer was cut or was not the data asked for; a check rejected the answer and the steps after it were skipped; the text is fluent but wrong — too long, a fact too many, a word in the wrong language; a figure is written the wrong way; or the price is not the estimate. Find the symptom below; each links to the page that explains the mechanism.

The run tells you where to look: every step reports its status, its error code and a message that, for these steps, says what to change (get_execution and get_execution_trace over MCP, GET /api/v1/executions/{id}, the step’s panel in the editor). A failed AI step is charged only what the provider billed for it: usually nothing for a provider error, the tokens used for a cut or invalid answer.

A model step failed

Code	What happened	What to do
AI_RESPONSE_INCOMPLETE	the model reached Max Tokens before finishing; reasoning counts too	raise Max Tokens, or lower the reasoning effort (Max Tokens and reasoning)
INVALID_STRUCTURED_RESPONSE	JSON mode, and the answer was not valid JSON	describe the object key by key in the prompt (Structured answers)
STRUCTURED_RESPONSE_NOT_OBJECT	JSON mode, and the answer was a list or a single value	ask for one object: {"items": [...]}
CATEGORY_NOT_IN_LIST	Image Classifier: the model answered a category that is not in your list	add other; say what each category covers (Image Classifier)
MISSING_PROMPT	no prompt on the node and nothing on its prompt input	write it, or check the step that should send it
EMPTY_RESULT	the model answered with nothing; the message says why when the provider does	rephrase, retry, or try another model
DOCUMENT_LIMIT_EXCEEDED, TOO_MANY_DOCUMENTS	Document Analysis: more than the provider accepts, or more than 20 documents	split the documents across steps (Document Analysis)
a provider’s refusal	the model declined the request	read the message; rephrase the prompt

A text step failed

Code	Node	What happened	What to do
FORMAT_TYPE_MISMATCH	Format as Text	a text where a number was needed, or a date not in ISO form (03/10/2026)	compute the number with Calculate; keep dates as 2026-10-03 (Format as Text)
INVALID_TEXT_FORMAT	Format as Text	an unknown language, currency or date style	use codes such as it-IT, EUR, long
TEMPLATE_PLACEHOLDER_NOT_TEXT, JOIN_INPUT_NOT_TEXT, REPLACE_INPUT_NOT_TEXT	Template, Join, Replace	a picture, a video or another file reached a text input	connect a text
INVALID_REGEX, REGEX_TIMEOUT	Replace Text	the pattern is not valid, or ran for more than 2 seconds	fix the pattern; avoid nested repetitions such as (a+)+ (Replace Text)
EXTRACT_TYPE_MISMATCH	Extract	a field of an AI answer has another type: "true" for true, or null	ask the model for the right type, and to leave out keys with no value (Structured answers)
JSON_PATH_NOT_FOUND	Extract	the model left out a key the workflow reads	require it in the schema, or give Extract a default

The steps after an AI answer were skipped

The run completed, but the outputs that depend on an AI answer are absent and their steps show as skipped.

• A schema check rejected the answer. Validate JSON Schema in mode route rejects an answer that breaks the contract — a missing key, a value outside an enum, a text over maxLength — and every step that reads it is skipped at no cost. Its errors output says which field broke which rule: tighten the prompt on that field, or relax the schema if the answer was right (Structured answers).
• A decision cut the branch: a Filter, a Select by Key with no match, an Extract with When Missing absent — see Branches and absence.

The text is fluent, but wrong

Symptom	Cause	What to do
Longer than asked	a length in a prompt is a request	check it with maxLength in a schema; never shorten with Max Tokens, which cuts and fails
A fact that is not in the data	the model filled a gap	give the facts and say they are the only ones; forbid guessing; for listings, take facts from Image Analysis, not from Product Description (Reading pictures)
A different text on every run	a model’s answer is a proposal	for a text that must not vary, use a template
The wrong language, or two languages mixed	the language was not stated, or a date or price was left in the source text	state the language; keep data out of the text and fill it per language (Languages)
A [marker] lost or translated	the model rewrote it	ask to keep every marker exactly as written; count them with Replace before filling
A sentence that does not agree with a value filled in	the translator did not know what the marker holds	say what each marker holds; choose a form that fits every language
[name] printed in the result	a marker with no port of that name	check the spelling of the marker and the port
null or an empty gap in a sentence	a value was missing	write the sentence so it reads without it, or build that part with Join Text (Text Template)
A prompt for a picture that misses the art direction	the style lives in each item’s prompt	put it once in the system prompt of the step that writes the prompt (Prompts for other AI steps)
A Video Analysis time a few seconds off	times from a video model are approximate	use Speech to Text captions for exact times
A figure in a document read wrongly	the model computed instead of copying	ask it to copy as printed; recompute with Calculate (Document Analysis)
Temperature seems to change nothing	the model does not use it (GPT-5.6, GPT-5.5, the o-series, Claude Opus 4.8 and 5)	say in the system prompt how free or steady the text should be

A figure is written the wrong way

Symptom	Cause	What to do
83.07 instead of 83,07 €	the number reached the template as it is stored	pass it through Format as Text with the reader’s locale and currency
A date as 2026-10-03 in a sentence	the same	Format as Text with date and a date_style
30% where German writes 30 %	the locale is not the reader’s	connect the reader’s locale to Format as Text’s locale input
A JSON document quoted as one long string	it was formatted as text	use Format as Text with json, which writes the document a text holds

The price is not the estimate

Symptom	Cause	What to do
The charge is far below the estimate	for text the estimate is a ceiling: Max Tokens written in full plus a reasoning allowance	nothing to fix; lower Max Tokens if the hold matters (What text costs)
A Video Analysis step cost far more than estimated	the estimate does not count the video’s length yet	budget about 0.31 credits per minute (Gemini 3.8 Flash); trim the video first
A step costs more than its neighbours	a model that reasons at length, or a long document or video read	lower the reasoning effort; send only the pages or minutes that matter (Keeping text cheap)