Skip to main content

AI Agent

The node that actually performs the AI-based analysis: it reads the document (or a segment) and produces comments according to the configuration of the chosen agent. It has five modes:

  • solo: it works alone;
  • supervised: it becomes a subagent driven by a Supervisor;
  • chat: it converses with the user;
  • form filling: it fills a PDF form (see Form filling);
  • consultable expert: it answers another agent's questions (see Expert agent).

Ports​

  • "Index" input (top, document index type) — present except in form filling and expert modes; connect it from a Document Index node if needed.
  • "Skills", "RAG" and "MCP" inputs (bottom) — always present too; connect them from a Skill node, a RAG node (knowledge base) and an MCP connector node respectively.
  • In solo mode, outside a Parallel container: a content input (left) and an annotation output (right), like a regular analysis node.
  • In supervised mode, outside a Parallel container: neither a content input nor an annotation output — instead, a dedicated "supervisor" output to connect to the "agents" input of a Supervisor node.
  • In chat mode: a content input (left, it receives the user's message), a dedicated "reply" output (right) to connect to the Chat output node, and a fifth MEM handle (bottom) for a Chat memory node.
  • "Experts" input (bottom, expert type) — in solo, chat, form filling and expert modes: each expert agent wired to it becomes a tool the agent can call to ask it a question.
  • In form filling mode: a content input (left, the PDF form) and a dedicated "form" output (right), to connect to the PDF form output.
  • In expert mode: no content input and no annotation output — a "consult" output (top), to connect to the "experts" input of the agent that may consult it.
  • Dragged inside a Parallel container: the content/annotation/supervisor ports are hidden (the container carries the group's port) — only the "index" input stays visible.

Node-level configuration (specific to this pipeline)​

  • Mode:
    • Solo: the agent analyzes directly within the pipeline's flow;
    • Supervised: it becomes a subagent, its document and annotation ports are hidden;
    • Chat: it answers the user's messages in plain text, with the conversation history in its context (no chunking, no verbatim anchoring). The History messages setting (20 by default) says how many previous messages are sent back at each turn;
    • Form filling: see the Form filling page;
    • Consultable expert: see the Expert agent page.
  • Agent — selection of the reusable agent to use (required); a button lets you create a new one directly from the panel.
  • AI model — optional override, for this node only, of the agent's default model.
  • Use support documents — unchecked by default, except in form filling mode; injects the text of support documents into this agent's context.
  • Role / objective / constraints overrides — locally replace (for this node/pipeline only) the chosen agent's defaults, without changing them globally.
  • Full prompt (advanced, hidden by default) — a raw override of the entire prompt.

Shared agent configuration​

The "Configure agent" button opens the configuration of the Agent entity itself — these changes apply to every pipeline that uses this agent, not just the current node. See AI Agents for details on the fields (role/objective/constraints, default model, knowledge base, MCP connectors, extraction passes…).

The MCP connector checkboxes shown directly in the node's panel also act immediately on the shared agent, not on data specific to the node.