Studio Chat

Studio Chat is built on LibreChat, an open-source chat interface, deployed on Microsoft Azure. Model requests are routed through LiteLLM, an internal proxy that standardizes connections to all underlying model providers. Usage and operational telemetry are captured via Langfuse for observability, tracing and cost management.

The chat models are served primarily from Azure AI Foundry. The image generation model (Flux.2 Pro) is provided by Black Forest Labs.

There are three key differences:

1. Authentication and governance. Access is gated by SMU single sign-on. Eligibility, quotas, and data policies are managed by the SMU AI Team (AIT).
2. Data handling. Prompts and outputs are not retained or used for training by the model providers. Conversation data is logged within SMU's environment for observability purposes only.
3. Choice of models in one place. You can switch between Anthropic, OpenAI, and DeepSeek models in a single interface without managing multiple accounts or subscriptions.

No. Studio Chat is a separate service. Copilot and other Microsoft 365 AI tools are integrated into specific applications (Word, Excel, Teams, etc.). Studio Chat is a standalone chat interface offering a broader range of models for general-purpose use.

If you believe you should have access but encounter issues (such as a redirect loop or an "Authentication Failed" message), contact us via the contact form. Please include your SMU email and a brief description of the error.

To request access, please use the contact form. We will review and respond to access requests according to the pilot’s eligibility criteria.

No. Access is tied to your individual SMU email account, and account sharing is not permitted. Your conversations, memory, and quota are unique to your account.

The pilot supports modern web browsers on both desktop and mobile devices. However, there is no dedicated mobile app available at this time. Access the platform using any up-to-date browser including Chrome, Edge, Safari, or Firefox.

There is no single "best" model. Consider these guidelines for selecting a model:

Experiment with multiple models for important tasks, as different models may offer valuable variations.

Yes. Toggle Web Search on in the input bar. The model will use web search results to inform its response and provide citations. Without this toggle, models respond only from their training data and any files you have provided.

Yes. The platform supports document upload (PDF, DOCX, DOC, TXT, MD, JSON), spreadsheets (XLSX, XLS, CSV), images (JPEG, PNG), and slide decks (PPTX, PPT). For longer documents, retrieval-augmented generation (RAG) is used to index content so models can answer targeted questions about specific sections.

See Files, Memory & Personalisation below for more.

Large language models include an element of statistical randomness when generating responses. The same prompt may produce slightly different answers each time, even with the same model. Different models will also vary more substantially. This is normal behaviour, not an error.

Yes. All large language models can produce content that is factually incorrect, outdated, or fabricated (commonly called "hallucinations") while sounding confident. Always verify any output you intend to rely on, share, or publish.

The pilot uses standard system configurations for each model. Custom agents, system prompts at the user level, Model Context Protocol (MCP) integrations, and code execution are not enabled in the pilot release but are on the roadmap (see Roadmap below).

No. Although image-to-image generation is supported, Flux.2 Pro is not very capable at specific edits to images as compared to GPT series of image model.
To generate an image with a reference image, upload your source image and provide instructions describing how you want your end result to be.

Yes. The image generation model has provider-side content filters. Prompts that violate those filters (for example, depictions of real people, violent imagery, or sexually explicit content) will not return an image. You must also comply with SMU policies, including the data classification rules — do not upload images containing sensitive, highly sensitive or personal data.

The pilot is intended for SMU work use. For external publication or commercial use, you should review the licensing terms of the underlying model and, where relevant, follow SMU's communications and branding guidance. When in doubt, reach out using this contact form.

We recommend disclosing AI-generated images when they are used in public-facing materials.

Yes, large files may be rejected or chunked automatically. If you encounter upload errors, try splitting your file into smaller pieces.
Currently the max file size limit per file is 10mb. You can upload a max total of 5 files at one time per prompt.

Memory allows the assistant to retain context across separate conversations, for example, your role, preferred writing style, or an ongoing project, so you don't have to repeat the same information each time. Memory is per-user and not shared with other users. You can view, edit, or delete individual memory items at any time from Settings → Personalization. You can also disable Memory entirely.

Saved Prompts let you save useful prompts for later reuse.

Bookmarks let you flag important conversations so you can find them quickly later. They are a simple organisation tool, your bookmarks are visible only to you.

Multi-Conversation lets you open two chat conversations side by side. This is useful for comparing how different models respond to the same prompt, or for working on two related tasks simultaneously.

You must not enter the following into prompts, attachments, or any other input 

• Sensitive data (high risk)
• Highly Sensitivedata (highest risk)
• Personal data under the Personal Data Protection Act (PDPA), such as names tied to identifiers, NRIC/FIN, contact details, addresses, financial details of identified individuals
• Student records, grades, transcripts, admission decisions
• Health or medical information
• Information received under non-disclosure or confidentiality agreements
• Embargoed research data
• Any other data classified above Restricted

If you are unsure, please reach out to us using this contact form.

No. Studio Chat is served via Microsoft Azure AI Foundry and other commercial endpoints under agreements where input and output data are not used to train or fine-tune the underlying models. SMU does not perform any custom training on your data

Yes. Each eligible user receives a monthly quota of 50 million LibreChat credits. Credits are consumed based on input length, output length, and the per-token cost of the model used. 

You can check your remaining balance under Settings → Balance.

Different models have different per-token costs from their providers. Generally:

• Larger, reasoning-heavy models (Claude Opus 4.5, GPT-5.4, DeepSeek-R1) consume more credits per response.
• Smaller, faster models (Claude Sonnet 4.5, GPT-5.2, DeepSeek-V3.2) consume fewer.
• Long inputs (large attachments, long conversations) and long outputs both increase credit usage. 
• Web Search consume more credits per response.

If you are running short on credits, switch to a lighter model for routine tasks.

Rate limits follows the compute capacity of the model as provided by Azure AI Foundry in the pilot, but fair-use limits apply. Excessive or automated usage may be flagged and throttled. Specific fair-use thresholds will be published as the pilot progresses after findings.

No. The pilot is free of charge to eligible users for in-pilot usage within the published quota. Cost recovery models may be considered if and when the service moves out of pilot.

Check that:

• The file is in a supported format (PDF, DOCX, DOC, TXT, MD, JSON, XLSX, XLS, CSV, PPTX, PPT, PNG, JPEG)
• The file is not encrypted or password-protected
• The file is not unusually large with many pages. If so, try splitting it into smaller files

The model may have hit its per-response output limit, the model rate limit or the conversation may have reached the model's context window. Try asking the model to continue, or start a fresh conversation summarizing the earlier discussion. You can also click on the regenerate button to try again.

Try regenerating the response, switching to a different model, or asking the model explicitly to format its output (for example: "Format the response in Markdown with clear headings.").

Make sure you have selected the correct model and have not accidentally included irrelevant attachments. Starting a fresh conversation often resolves drift in long chats. Keep chat sessions well segregated by relevant context and do not mix too many different task in one chat session.

Reach out to us using this contact form. Keep a screenshot of the error, take note of the task you are doing and the model in use. This will help us to identify the root cause better when we do a follow up on the issue you encountered.

Use the contact form to submit suggestions. We welcome input from the SMU community and use it to prioritise roadmap items.

Please reach out to us through the contact form here.

The platform is operated by the SMU AI Team (AIT), with support from IITS.