AI Assistant

The AI Assistant is a multi-session workbench for asking questions about the pipeline or workflow you have open. Sessions stay available while you move between files. Open it as:

It is off by default. Nothing is sent to a language model until you enable it and create an AI Provider.

See AI technology for providers, privacy, and how other plugins add assistants.

Create at least one named provider in the Metadata perspective (AI Provider) and use Test on the editor to confirm the endpoint responds. Put the API key in an environment variable or a variable resolver, not as a literal in the metadata JSON. See Secrets and personal information.

AI-applied edits can change pipelines, workflows and metadata. Graph Undo only reverses some canvas edits. Metadata saves and pasted transforms or actions are not undone that way.

Before you apply proposals:

The assistant shows this reminder on an empty session, in Configuration → Plugins → AI Assistant, and on the review dialog above Apply selected.

From a pipeline or workflow graph:

Those actions open a floating window so the pipeline or workflow stays on screen. If the assistant is already docked or already floating, that layout is reused.

The workbench reuses the session of that pipeline or workflow when there is one, also after you saved a new file for the first time, or closed and reopened it. Two files with the same name in different folders each get their own session.

New session (+) starts a session for the pipeline or workflow that is open in the Explorer perspective. With no file open it starts a session that is not linked to one; the pipeline and workflow advisors need a file, so open one and click Link to the open file on the session toolbar. CTRL-SHIFT-A (macOS: Cmd-Shift-A) on an open pipeline or workflow does what AI Assistant on its toolbar does. Elsewhere it opens the perspective.

Sessions belong to the project they were started in: switching projects shows that project’s sessions. With Keep conversations after Hop GUI closes (Configuration → Plugins → AI Assistant, on by default) they are saved per project in the audit folder, ${HOP_AUDIT_FOLDER}, where Hop GUI also remembers the open files of each project. That folder is on your machine and outside the project, so conversations do not end up in version control. Questions, answers, proposals and the session options are saved as plain text; the context sent with a question is not. What you type in a question is saved as you typed it, so do not paste secrets or personal data into questions, or turn the option off. When you close a pipeline or workflow, its session lets go of it; open the file again to continue the conversation. Closing a session that has a conversation asks for confirmation first, and stops a question that is still running. Closing the window or dock tab during a question lets it finish in the background, unless When a window closes during a question is set to Cancel the question.

Each session has:

Every question sends the context that is checked at that moment, including follow-up questions. After the pipeline or workflow has run, Include execution log excerpt switches on by itself, visibly in the Sharing line, and the log of the latest run is sent. When you switch it off, that holds until the next run: a new run brings a new log, and the option switches on again. The previous questions and answers of the session are sent as conversation history.

The advisor plugin decides what can be shared. By default little extra is shared — the collapsed Sharing line is the short version of the current send. The plugin catalog is on by default so apply-able proposals can use real plugin ids.

Click the line (or the arrow) to expand the checkboxes. Every checked inclusion is copied into the prompt that goes to the language-model service. That has two costs: tokens (money and latency) and exposure (anything in the prompt leaves your machine). Turn on only what the current question needs, then collapse the panel.

Hover a checkbox for a short reminder of cost and risk. The collapsed Sharing line lists what you chose first and sums up the items that always go along as the basics; hover it for the full list with a short explanation of each. Show what is sent… shows the exact text.

Before a question is sent, the assistant estimates its size: the instructions, the checked context, the conversation history and room for the answer. When that clearly exceeds the Context size of the AI Provider, the question is not sent and the transcript explains what to uncheck. This matters most for local Ollama models, which would otherwise drop the start of the prompt without a word. For a hosted provider without a Context size, the question is checked against 200000 tokens for Anthropic and 128000 for the others.

The assistant answers in the language of your question.

The assistant is only as safe as the project you point it at. If a password, token or personal record is hard-coded in a pipeline, workflow, metadata object, note, SQL statement or Data Grid sample, it can appear in XML, check results or logs and be sent to the AI service.

Set this up before you enable the AI Assistant:

Before anything is sent, the assistant masks:

A value that only references a variable, such as ${DB_PASSWORD}, is not a secret and is sent as is, so the model can see which variable holds it.

Redaction does not catch literals in SQL, notes, free-text logs, or a plain (unencrypted) secret in a field with an ordinary name. Host names and user names are sent. The checkboxes above are the real control.

The AI Provider API key is never copied into advisor prompts. It is still project metadata if you paste it: use ${AI_API_KEY} or a resolver there too.

The same workbench can run outside the perspective, like the Database perspective and Search:

Hop GUI remembers where you put it: after Dock, AI Assistant on a toolbar opens it in the dock, and the dock tab is opened again when Hop GUI starts. Float and Dock move the assistant: moving it from the dock to a floating window closes the dock tab, and the other way around. The button of the place you are in is disabled. The assistant is open in one place at a time. Float or Dock in the perspective switches back to the files, opening the perspective closes the floating window and the dock tab, and Tools → AI Assistant window moves the assistant from the dock to a floating window, as Float does.

The example above is a typical error-diagnosis session: the workflow graph and Logging tab stay visible while the assistant explains that the failure is a missing execution-information cache folder, not a broken hop.

When you ask for a change (build, add, connect, fix, create …), the model appends proposals to its answer. Questions that only ask to explain or diagnose get a written answer without proposals. The transcript then shows Review N proposed change(s). When the proposals in an answer cannot be read, or do not follow the proposal schema (a known type, a risk level of LOW, MEDIUM or HIGH, and parameters), the assistant asks the model once to correct them; if that fails too, the transcript says why. With Structured answers on the AI Provider, the provider holds the model to that schema, so the proposals cannot come back broken.

Supported pipeline types: add/delete/rename transform, add/delete hop, set location, add note, copy transform XML to the clipboard, replace an existing transform from XML. Supported workflow types: add/delete/rename action, add/delete hop, set location, add note, copy action XML to the clipboard, replace an existing action from XML. Both sessions also accept copy-metadata-JSON and save-metadata proposals.

ADD_TRANSFORM and ADD_ACTION load the plugin with defaults only. Field mappings, SQL and connection names are not applied that way.

To change an existing node’s configuration, apply REPLACE_TRANSFORM / REPLACE_ACTION (plugin type must match; name and canvas location stay). To add a fully configured node, apply CLIPBOARD_TRANSFORMS / CLIPBOARD_ACTIONS and paste on the canvas. SAVE_METADATA creates or overwrites a named metadata object; it is not undone with graph Undo but with Undo metadata changes in the transcript. Proposals can only save the metadata a user typically asks the assistant to create: relational database connections, Cassandra, MongoDB and Neo4j connections, Neo4j graph models, data sets, unit tests, file definitions, schema definitions and partition schemas. Other metadata can decide where prompts, secrets, files, logs or execution go (AI providers, variable resolvers, run configurations, Hop servers, logging, web services, VFS and Git connections, …). Create or change it yourself in the Metadata perspective.