Skip to main content

AI Assistant — Example Interactions

Experimental

The AI Assistant is an experimental feature and is under active development.

This page shows real examples of what you can do with the HPCBOX AI Assistant — both from the built-in chat panel in ClusterApp and from an external MCP client such as Claude Code or Claude Desktop. Each example shows the full conversation and the outcome.

Tagging Output Files and Recalling Them Later​

One of the most useful things the assistant can remember is where your simulation results live. Instead of keeping notes in a spreadsheet, you can tag output files once and ask the assistant about them in any future session.

Tagging Files​

User:

I would like to remember these output files and associate them with the turbine project.

Assistant:

Sure. Let me grab the current workflow details first so I can include the right context.

(reads workflow — ANSYS CFX, 256 cores, project TurbineStudy01, working directory /home/alice/turbine)

Got it. Please select the output files you want to remember.

[Browse Files button appears in the chat]

The user clicks Browse Files, selects the result files from the file dialog (e.g. turbine_001.res, turbine_001.out, post_results.cdat), and confirms.

Assistant:

3 files selected. Saving to memory...

Done. I've saved the following:

"ANSYS CFX → CFD-Post workflow for project TurbineStudy01 in /home/alice/turbine, 256 CPU cores. Output files: turbine_001.res, turbine_001.out, post_results.cdat."

You can ask me about these files in any future session, even after closing the app.

What Happened​

Behind the scenes the assistant:

  1. Called GetWorkflowDetails to read the open workflow's application, settings, and project ID
  2. Sent a RequestFilePaths request which displayed the Browse Files button in the chat
  3. Stored the selected file paths as a reference token in the local database
  4. Called SaveMemory with a summary combining the workflow context and the file reference token

The file paths are stored locally. If Private file references is enabled in settings, paths are stored as opaque tokens and never sent to the AI provider.

Recalling the Files Later​

In a later session (or after restarting the app):

User:

Where are the output files for the turbine study?

Assistant:

(searches memory)

Found it. For project TurbineStudy01 (ANSYS CFX → CFD-Post, 256 cores, /home/alice/turbine), the output files are:

  • turbine_001.res
  • turbine_001.out
  • post_results.cdat

These were saved after the last run. Would you like me to do anything with them?

Creating a Workflow from a Description​

User:

Create an OpenFOAM workflow for project Intake01. I want to mesh with blockMesh first and then run simpleFoam. Use 64 CPU cores. Working directory is /home/alice/intake.

Assistant:

Creating a 2-step OpenFOAM workflow (blockMesh → simpleFoam, 64 cores, /home/alice/intake)...

(lists applications, gets OpenFOAM details, validates, creates workflow)

Done. The workflow has appeared in a new editor tab. I've also saved a memory of this for future reference.

What was saved to memory:

"Created OpenFOAM workflow (blockMesh → simpleFoam) for project Intake01 in /home/alice/intake. 64 CPU cores."

In a later session:

User:

What settings did I use for the intake project?

Assistant:

(searches memory)

For project Intake01: OpenFOAM with blockMesh → simpleFoam, 64 CPU cores, working directory /home/alice/intake.

Diagnosing a Failed Job​

After a workflow runs and a step fails, the assistant can read the log files and identify the root cause.

User:

The CFX step failed. What went wrong?

Assistant:

(searches memory for CFX log file guidance) (reads step log — hpcbox/CFX-Solver-1.log.2)

The CFX solver failed with:

ERROR: Unable to checkout license for feature aa_r_cfd.
License server reports: No licenses available.

This is a license availability issue — all CFX licenses are currently in use. You can:

  • Retry later when a license frees up
  • Check current usage with lmstat -a -c @license-server
  • Ask your admin whether additional licenses are available

If the scheduler log doesn't contain the full error (common with applications like ANSYS CFX where detailed errors go to a .res file), the assistant scans the working directory:

Assistant:

The scheduler log only shows an exit code. I found these files in the working directory that may contain more detail:

  • turbine_001.res — 42 KB, modified 3 minutes ago
  • turbine_001.out — 8 KB, modified 3 minutes ago

Shall I read these to look for errors?

User:

Yes.

Assistant:

(reads turbine_001.res)

Found the error in turbine_001.res:

ERROR: Mesh topology check failed — non-conformal interface detected at boundary "inlet_wall".

This is a mesh issue, not a license or resource problem. The inlet/wall interface geometry needs to be checked in your meshing tool.

This application's errors appear in the .res file rather than the scheduler log. Should I save this as shared knowledge for the team, so future CFX failures go straight to the right file?

User:

Yes.

Assistant:

Saved. Next time a CFX step fails, I'll check the .res file directly without needing to scan.

Non-Standard Log Extension​

Some in-house or third-party applications write their errors to files with non-standard extensions. The agent handles this too.

User:

The solver step failed. Can you find out what went wrong?

Assistant:

(searches memory — no guidance found for this application) (reads scheduler log — shows only exit code 1)

The scheduler log only shows that the step exited with an error. I didn't find any .log, .res, .dat, .msg, or .out files in /home/alice/myjob.

Would you like to browse for the log file?

User:

Yes, there should be a .hpcerr file in there.

Assistant:

(scans for .hpcerr files — finds solver.hpcerr)

Found solver.hpcerr (12 KB, modified 2 minutes ago). Shall I read it?

User:

Yes.

Assistant:

(reads solver.hpcerr)

Found the error:

FATAL: Input deck parse error at line 47 — unknown keyword "BOUNDRY" (did you mean "BOUNDARY"?)

There is a typo in your input file at line 47.

This application writes its errors to .hpcerr files rather than the standard log locations. Should I save this so the team knows where to look next time?

User:

Yes.

Assistant:

Saved: "For [AppName] jobs, error output is written to .hpcerr files in the working directory. Check <working_dir>/*.hpcerr for failures."

Next time a step for this application fails, I'll scan for .hpcerr files immediately.

Asking About a Colleague's Work​

When shared context is enabled, the assistant can search across all team members' saved memories.

User:

Has anyone on the team run an OpenFOAM simulation for the intake manifold project in the last 30 days?

Assistant:

(searches shared memory)

Yes — Bob ran an OpenFOAM workflow (blockMesh → simpleFoam) for project IntakeManifold02 on 2026-03-28 in /projects/manifold, using 128 CPU cores.

If no results are found, the assistant reports clearly:

Assistant:

No results found in shared memory for OpenFOAM/intake manifold in the last 30 days. The team may not have run one recently, or shared context may not be enabled for all users.

MCP Server Examples (Claude Code / Claude Desktop)​

These examples show the same assistant capabilities accessed from an external MCP client, such as Claude Code in VS Code or Claude Desktop. The workflow is the same — the client's own LLM calls HPCBOX tools to build, run, and query workflows.

Create and Execute a Workflow​

User: Read the OpenFOAM tutorial at [url] and create an HPCBOX workflow for it,
      then execute it on the cluster.

Claude: [fetches tutorial, identifies steps: blockMesh → decomposePar → simpleFoam → reconstructPar]
        [calls ListApplications, GetApplicationDetails("OpenFOAM")]
        [calls ValidateWorkflow, CreateWorkflow]
        Workflow created: /home/alice/workflows/openfoam-cavity.hpcbox

        Shall I execute this now? It will run 4 steps using 32 CPU cores in /home/alice/cavity.

User: Yes.

Claude: [calls ExecuteWorkflow]
        Workflow started. Steps are being submitted to the scheduler.
        [calls GetExecutionStatus — polling]
        All 4 steps completed successfully in 23 minutes.

Search Team Memory​

User: What OpenFOAM settings has the team used for external aerodynamics cases?

Claude: [calls SearchMemory("OpenFOAM external aerodynamics")]
        Found 3 relevant entries:
        - Bob (2026-03-10): snappyHexMesh → simpleFoam for project DragStudy01,
          256 cores, working dir /projects/aero
        - Alice (2026-02-28): blockMesh → rhoPimpleFoam for project SupersonicInlet,
          128 cores, working dir /projects/supersonic
        - Charlie (2026-01-15): cfMesh → pimpleFoam for project WingProfile03,
          64 cores, working dir /projects/wing

Diagnose a Failed Job​

User: The last workflow I ran failed. What went wrong?

Claude: [calls SearchWorkflows — finds most recent workflow]
        [calls GetExecutionStatus — step "simpleFoam" failed]
        [calls SearchMemory("OpenFOAM simpleFoam error log location")]
        [calls ReadJobLogs — reads hpcbox/simpleFoam.log.1]

        simpleFoam failed with:
          "Floating point exception (core dumped)"
          Last residuals: Ux=8.2e+02, Uy=9.1e+02 (diverging)

        This is a solver divergence — the case is numerically unstable.
        Common causes: time step too large, poor mesh quality near boundaries,
        or relaxation factors set too high. Check your fvSolution and fvSchemes.

Tips​

  • Memory is automatic — the assistant saves a summary after every workflow it creates or modifies. You don't need to ask.
  • Be specific when recalling — include the project name, application, or approximate date for faster results. "What were the CFX settings for TurbineStudy01?" works better than "what did I do last time?"
  • Shared context requires setup — ask your administrator for the shared context directory path if you want to search across team members' memories. See Agent Shared Context.
  • Private file references — if your site requires that file paths not be sent to the AI provider, enable Private file references in the agent settings panel. File associations still work; paths are just resolved locally after the AI responds.