Skip to main content
Components (sometimes called Processors or Nodes) are the fundamental executable blocks inside a Workflow Definition. Each component adheres to a specific Interface (e.g., genai.chat) which dictates its input ports, output ports, routing logic, and functional behavior. Below is an overview of the supported taxonomy of component interfaces natively powered by BlobHub.

Component Taxonomy

Flow Control (flow.*)

These handle the fundamental initiation and termination of the logic loop.
  • flow.start: The absolute entry point of the execution.
  • flow.end: A terminal endpoint resolving the execution gracefully.
  • flow.goto: A redirect node pointing to another discrete location.

Data Variables (data.*)

Static blocks providing statically typed constants injected dynamically into the runtime graph.
  • data.integer: Hardcodes a number.
  • data.message: Hardcodes a rich semantic AI message prompt.
  • data.messages: Hardcodes conversation history arrays.

Logic (logic.*)

Control branching, data manipulation, and programming logic.
  • logic.code: Escapes the interface into raw Python sandboxes (See Code Component).
  • logic.generic_provider: Injects generic authenticated API blocks.
  • logic.for: Loops over input arrays.
  • logic.sleep: Pauses execution dynamically.
  • logic.insert: Injects dynamic text strings into payload templates.

Generative AI Models (genai.model)

A provider for every model vendor, which the chat and agent components call.
  • genai.model: A model of any vendor in the catalog, with a credential per vendor. A run may name another. (See Model Provider.)

Generative AI Chat (genai.chat)

One model call over a conversation.
  • genai.chat: Multi-turn messaging based on Generative AI. It takes its model from a genai.model through its model port. (See Chat Component.)

Generative AI Agents (genai.agent.*)

Recursive execution loops allowing LLMs to invoke tools autonomously.
  • genai.agent.basic: Implements a standard reasoning and execution (ReAct) agent looping wrapper mapping tools natively to workflow functions.
It takes a genai.model on each of generic_model, planning_model, tool_use_model, tool_response_model and response_model, so each step may use another model. A tool that answers HTTP 400 or above posts one processor.warning, Tool <name> answered HTTP <status>, and the run goes on: the answer’s body still becomes the step’s observation. A GET or POST to an https:// tool that still answers 502, 503 or 504 after three retries fails the call instead.

Vector Database (database.vector.*)

Semantic search integrations natively wired to external retrieval-augmented generation (RAG) providers.
  • database.vector.pinecone: Direct Pinecone ingestion and search, with an api_key credential.
  • database.vector.ingest: Abstract interface pushing embeddings.
  • database.vector.retrieve: Abstract interface pulling similarity searches.

The Web (web.*)

Fetching, scraping, and dynamically parsing the actual internet.
  • web.scrape: General HTML fetching.
  • web.scrape.firecrawl: Sophisticated headless browser scraping via Firecrawl, with an api_key credential.
  • web.search: General SERP queries.
  • web.search.tavily: Web synthesis and targeted searching via Tavily, with an api_key credential.

Text Processing (text.*)

Splitting and parsing text.
  • text.chunk: Breaks down giant documents into smaller token-ready vectors for database storage.
  • text.render_template: Parses Jinja templates with dynamic arguments.