Arch AI – Help and FAQ

Initial version: March 25, 2024 – Last update: October 16, 2026

What’s new

October 2026: New: the Code Interpreter. Agents can now write and run code, and every code block in a chat has a “Run Code” button. Everything runs in a sandbox on D-ARCH servers, see below.

October 2026: We have changed the login backend. A few users may find that their chats and files are no longer displayed as usual. In this case, contact us and we will resolve the issue.

Overview

Arch AI (archai.ethz.ch) is the AI platform of the Department of Architecture, available to all members of ETH Zurich. It gives you chat with large language models, work with your own documents, image generation, voice input and output, web search, and a builder for your own assistants (“agents”).

One of Arch AI’s main objectives is to provide a straightforward user interface for on-premises AI. The aim is to highlight the capabilities and limitations of on-premises AI running on off-the-shelf hardware.

Everything runs on servers of the Department of Architecture at ETH Zurich. The only exception is one commercial cloud model, which you have to pick explicitly.

Where your data is processed

Chat history is stored in a database at D-ARCH and is visible only to you. Members of IT Services D-ARCH can access the database for operations and support.

Costs

Local services are free of charge and will remain so. The use of GPT is paid for by IT Services D-ARCH. We reserve the right to remove cloud models if costs are too high. If your unit wants to rely on Arch AI permanently in teaching or research, contact us.

Getting started

Login

A valid ETH account is required. You log in with your ETH username and password through the ETH login page, followed by multi-factor authentication.

The interface

• Left sidebar: your conversations, bookmarks, prompts and agents.
• Model selector (top left): choose between local models and GPT.
• Attach and tool selector menu next to the message box: attach files, enable web search, switch tools and memory.
• Microphone: dictate your message.

Choosing a model

Please note, available models change over time.

Chat with your own files

Attach PDF, Word, text or spreadsheet files to a message. Arch AI reads the file and answers based on its content. Large files are indexed for search, so you can ask questions across many pages or many documents. Files and the index stay on ETH servers.

Images attached to a message are described by the model (Gemma 4 and GPT understand images).

Web search

Switch on Web search in the “tools” menu. The model then searches the web, reads the pages it finds, and cites them in its answer. The search runs through our own search server, which forwards the search terms to public search engines. Your conversation is never sent to them, only the search query the model formulates.

Images

Generating images

Images are generated on D-ARCH servers. Two shared agents are ready to use, pick one in the Agents list:

Both are worth a try with the same prompt; pick the one whose results you like better.

Describe the image you want in plain language. A single image takes from a few seconds up to a minute. You can also add the image tool to your own agent, for example with a fixed style description for a studio. The images belong to you and may be used freely. We recommend marking AI-generated images as such.

Understanding images

Attach a photo, plan or screenshot and ask about it. Works with Gemma 4 and GPT.

Voice

• Dictate: click the microphone, speak, and the text appears in the message box. Transcription runs on an ETH server.
• Listen: every answer has a speaker icon that reads it aloud. Voices are available in German, French, Italian, English, Dutch, Spanish and Portuguese.
• Conversation mode (Settings > Speech): hands-free back and forth.

Prompts

Prompts are saved text templates. Instead of typing the same instruction every time, you save it once and call it with / in the message box. A prompt can contain variables in double curly braces, which you fill in when you use it. Prompts can be shared with other users.

Examples:

• Polish this abstract for an English-language journal, keep the technical terms: {{text}}
• Turn these meeting notes into a structured protocol with decisions and open tasks: {{notes}}
• Summarise the attached paper in five bullet points for a non-specialist audience.

Agents

An agent is a model plus instructions, optionally with files and tools. You build it once and reuse it, or share it with your group.

What is the difference?

Building an agent

Open Agents in the sidebar and click Create. Give it a name, choose a model, write the instructions, and add what it needs:

• Files for file search: upload documents the agent should know.
• Tools: web search, image generation, code interpreter, artifacts (interactive previews of code and documents).
• Share: with named people, with groups, or with everyone on Arch AI.

Examples

• Regulations assistant: upload the department’s regulations and guidelines as PDFs, instruct the agent to answer only from these documents and to cite the section. Share it with your institute.
• Research assistant: web search enabled, instructions to always list sources and to prefer academic publications.
• Image generator for studio work: duplicate the “Qwen-Image Agent” and add a fixed style description to the instructions, so every student in the studio gets consistent results.
• Code reviewer: Local Code AI, instructions to review Python scripts for readability and errors.
• Data assistant: Local Code AI with the code interpreter enabled, instructions to load attached CSV or Excel files with pandas, describe the data first and produce charts as PNG files.

Skills

A skill is a set of written instructions for one kind of task, for example a review checklist, writing guidelines or a standard research procedure. Unlike a prompt, you do not send it yourself: an agent picks up the skill on its own when your request matches its description.

• Open Skills in the sidebar and click + to write a skill (name, a short description of when it applies, and the instructions) or upload one as a .md, .zip or .skill file.
• In the agent builder, enable Skills and choose which skills the agent may use.
• In a chat, type $ to apply a skill to your next message by hand.

The description decides whether the agent uses the skill, so state clearly when it applies.

Memories

Memories are short facts about you that Arch AI adds as context to every conversation, for example your preferred language, your field of work or a current project. Memories are not extracted from your chats; you write them yourself.

• Open Memories in the sidebar to add, edit or delete entries. A few short entries work best.
• Switch memory on or off in the “tools” menu next to the message box. The setting applies to all your conversations.
• An agent with the Memory tool enabled can also save a fact when you ask it to (“remember that …”).
• Memories are stored at D-ARCH and are visible only to you.

Bookmarks

Tag conversations with bookmarks to find them again. Click the bookmark icon in a conversation, create a tag, and filter the sidebar by tag.

Code Interpreter

Arch AI can run code and give you the result: analyse a data file, draw a chart, convert a document, check a calculation. The code runs in an isolated sandbox on D-ARCH servers. Each run gets its own small virtual machine that is discarded afterwards, so nothing can affect other users or our systems. Languages: Python, JavaScript/TypeScript, Go, C/C++, Java, PHP, Rust, Fortran and R.

Two ways to use it:

• Through an agent: enable Code Interpreter in the agent’s tools. The agent writes and runs code on its own when the task calls for it, for example “Plot the monthly values in the attached CSV as a bar chart” or “Turn this table into an Excel file”. Files the code produces appear as attachments in the answer.
• Run Code button: every code block in a chat has a Run Code button that executes the snippet and shows its output.

Good to know:

• To let the code work on a file, simply attach it to your message with the paperclip. In a chat with an agent that has the code interpreter, the file is placed in the code’s working folder automatically, you never need the path yourself. Files the code saves there come back as downloads in the answer, up to 10 per run.
• Python comes with the usual libraries preinstalled: data and statistics (pandas, numpy, scipy, statsmodels, scikit-learn), charts (matplotlib, seaborn, plotly), Office and PDF files (openpyxl, python-docx, python-pptx, pypdf2, reportlab), images (pillow, opencv), and geodata (geopandas, rasterio, osmnx, folium).
• The sandbox has no internet access and no keyboard. Code cannot download packages or data, and input() returns an empty string. Put values directly into the code or ask the agent to use parameters.
• A run may take up to five minutes. Longer tasks are split by the agent into several runs.
• Code, files and results stay on D-ARCH servers.

The model wrote and ran this stair calculator on the first try. The table looks right, but the last column is off: a stair with n risers has n-1 treads, so the run is about 30 cm too long. Code that runs is not code that is correct. Check results you rely on.

API access

For research projects of the department or for projects with sensitive data, the local models are also available through an OpenAI-compatible API (HTTPS, streaming, tool calling).

The API speaks the OpenAI protocol, so any OpenAI client library works. Both examples use the official openai Python package (pip install openai). You receive the host, the token and the model name from us.

Ask a question

from openai import OpenAI

HOST = "https://on-premise-llm.ethz.ch/v1"
TOKEN = "your-api-token"
MODEL = "model-name"

client = OpenAI(base_url=HOST, api_key=TOKEN)

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", 
               "content": "What is the tallest building in Zurich?"}],
)
print(response.choices[0].message.content)

➜ python ask-a-question.py   
The tallest building in Zurich is the **Prime Tower**. 
It stands at approximately 126 meters (about 413 feet) tall.

Tag an image

The image is sent as part of the message; the model answers with a list of tags.

import base64
from openai import OpenAI

HOST = "https://on-premise-llm.ethz.ch/v1"
TOKEN = "your-api-token"
MODEL = "model-name"

client = OpenAI(base_url=HOST, api_key=TOKEN)

with open("photo.jpg", "rb") as f:
    image = base64.b64encode(f.read()).decode()

response = client.chat.completions.create(
    model=MODEL,
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", 
             "text": "Describe this image with 5 to 10 tags, \
                     comma-separated, nothing else."},
            {"type": "image_url", 
             "image_url": {"url": f"data:image/jpeg;base64,{image}"}},
        ],
    }],
)
print(response.choices[0].message.content)

➜ python tag-an-image.py
Zurich, Switzerland, Cityscape, Architecture, River, 
Lake, Alps, Church, Skyline, European

Request access

See IT ARCH - AI Tools and contact us if you require API access.

Contact

Suggestions and feedback are welcome.

IT Services Group Department of Architecture
servicedesk@arch.ethz.ch

ETH Zürich
IT Services D-ARCH
Thomas Gemperli
HIL C 38.2
Stefano-Franscini-Platz 5
8093 Zürich