Skip to main content

​
Overview

The text generator node produces LLM output using the Vercel AI SDK (generateText). The Worker loads node configuration from Postgres, resolves the concrete model via system_ai_models, builds system/messages/tools, calls the provider, persists results on the node row, and returns token usage to the workflow for billing. textGeneratorActivity delegates to TextGeneratorService.process. For product and UI details (inputs, provider knobs, model capability tables), see AI Text Generator.

​
When it runs

textGeneratorActivity is invoked from the processSingleNode workflow when the node type is text generator (NodeType.TEXT_GENERATOR). The workflow typically calls validateModelAccessActivity before execution and chargeTokensActivity afterward using the returned token counts.

​
Activity signature

async textGeneratorActivity({
  nodeId: string,
  sessionId: string,
  userId: string,
}): Promise<TextGeneratorResponse>
TextGeneratorResponse: 
{
  output?: string;
  inputTokens?: number;
  outputTokens?: number;
}
The activity returns only these fields; richer state (generated text, logs, previews, memory) is written to flows_nodes.data inside TextGeneratorService.process.

​
End-to-end flow

  1. Load node — SELECT from flows_nodes by nodeId; read data (prompt, provider fields, memory flags, etc.).
  2. Resolve model field — From data.AiSystem, the service picks which property holds the model id (aiModel, groqModel, geminiModel, anthropicModel, or perplexityType). See getModelFieldByAiSystem in the service.
  3. Load model metadata — SELECT from system_ai_models where modelName matches that id (aiCompany, limits, isVisionModel, etc.).
  4. Instantiate model — selectModel maps aiCompany to a client (openai, anthropic, gemini, perplexity, xai, groq) and returns a LanguageModel. Anthropic ids are normalized via an internal map when the stored id matches known keys.
  5. Build request — buildConfig composes system, messages, optional web-search tools, and cleanedPrompt (used for memory). Persona and instructions append to system. Prompt lines come from promptData + prompt; context from contentData + content. Memory injects prior turns when memory is true and sessionId is present.
  6. Generate — generateText with provider-specific options from buildGenerateTextOptions (temperature, max tokens, stop sequences, Perplexity search context, OpenAI reasoning effort, etc.). autoToken interacts with maxOutputToken from system_ai_models.
  7. Persist — UPDATE flows_nodes with text, structured logs from buildAILog, executionStatus: "COMPLETED", nodeUsedTokens, appended previewResponses, and updated userSessionsChat when memory is enabled.
  8. Return — output, inputTokens, outputTokens from result.totalUsage.

​
Persisted node data highlights

FieldRole
textModel output string.
logsHuman-readable execution log (model id, tools used, search queries, URLs, token usage, finish reason).
executionStatusSet to COMPLETED on success.
nodeUsedTokensInput token count from usage (used for display; billing uses activity return values).
previewResponsesPrevious outputs plus the new output appended.
userSessionsChatWhen memory is true, the current session’s memories gain new user/assistant entries from this run.
When node.webSearch is true, buildConfig attaches provider-specific tools:
AiSystemTool source (conceptually)
OpenAIopenai.tools.webSearch
Anthropicanthropic.tools.webSearch_20250305
Geminigoogle.tools.googleSearch
Other providers throw if web search is requested without support. Logs include tool calls, search queries, and URLs when present.

​
Multimodal

If multimodalEnabled is true and system_ai_models.isVisionModel is true, URLs in the prompt are extracted; non-file URLs are stripped from the text sent as prompt. File URLs are validated (type, size via HEAD), capped at 3 files, and merged into messages as image or file parts depending on provider. Limits include max sizes per type (images 5MB, generic 10MB, PDF/video/audio rules as in code).

​
Session memory

When memory is true, sessionId is required (throws if missing). The service loads userSessionsChat for that session, applies processMemoryFilter using selectedType / selectedNumber / selectedIndex (first/last N entries or words, or all), and prepends those messages. After generation, if cleanedPrompt is non-empty, new user and assistant memory entries are appended for that session.

​
Provider-specific generation options

buildGenerateTextOptions branches on modelInfo.aiCompany (lowercased). Typical mappings:
  • Anthropic — anthropicToken, anthropicTemperature, anthropicTopP, anthropicTopK; max output resolved with autoToken vs system_ai_models.maxOutputToken.
  • Gemini — geminiMaxOutputTokens, temperature, topP, topK.
  • OpenAI — aiTokens, aiTemperature, aiPp (top P), aiFp (mapped to topK in options), optional aiEffort → providerOptions.openai.reasoningEffort.
  • Perplexity — max_tokens / perplexityTokens, temperature, top P/K, search_context_size in provider options.
  • XAI — aiTokens, aiTemperature.
  • Groq — groqMaxTokens, temperature/topP (with fallbacks), seed, groqStop as stop sequences.
Unsupported or missing providers throw during model selection.