CometAI Known Bugs & Issues
Below is the list of issues identified by the CometAI Platform Team and reported by initial testers.
Important: If your issue is already listed, please do not email us or create a Service request, as we are aware of it and are actively working on a resolution. List will be updated as issues are resolved or as new issues arise. This Article has been updated post CometAI 3.0 Rollout in July 2026
- CometAI Conversations, Prompt Templates, Custom Instructions, Follow‑up Buttons Access across Multiple Devices:
- These are currently stored locally on the user’s browser/device, regardless of conversation storage or sync settings. As a result, these features do not automatically synchronize across different browsers or devices.
- If users want to reuse them when switching browsers or devices, they should save their prompts, conversations and instructions externally (for example, in a Word document) and recreate them as needed.
- This is a known issue, and the CometAI team is actively working to address it in a future release.
- Conversation Tagging: Users can currently add tags to conversations, but searching by tag name in the left-hand Conversations sidebar does not return tagged chats.
Tags are therefore not reliably retrievable from the LHS search bar at this time. This is a known issue; behavior is under investigation for a future fix.
- Import/Export Conversations: The function and purpose of the Import/Export Conversations feature are not clear to users yet and needs to be more properly documented after researching more. Hence, it is advised to not use this feature until further notice.
- LaTeX Rendering: Currently, the platform supports simple to moderate LaTeX type equation rendering. For complex LaTeX mathematical equations, it is advised to use the latest models such as GPT 5.5, GPT 5.4-MINI, Gemini-3.5-Flash, Claude Opus 4.8.
- Large File Uploads (Local Files, Assistants, and M365): When uploading larger files of approximately over the size of 10MB for PDFs , users may receive the following toaster notification:
"Please wait for a few minutes for file embedding to be completed. If you do not get a response, try again." In some cases, the platform may return an empty response even after waiting 10–15 minutes for the embedding process to complete.This is a known issue affecting large file processing and ingestion workflows. The CometAI team is actively investigating the behavior and working on a resolution in a future release very soon.
-
Llama Models – Custom Instructions & Image Upload Limitations:
-
Custom Instructions (CI) are not recommended with Llama models, as responses may be inconsistent or may not reliably follow the configured instructions. When using image inputs, upload no more than three (3) images per prompt. Uploading more than three images simultaneously may result in incomplete processing, degraded performance, or failed responses.
-
Mistral Models – Custom Instructions & Diagram Assistance:
-
Custom Instructions (CI) are currently not functioning reliably with Mistral models and may be ignored during response generation.
-
Diagram assistant feature is not supported consistently on Mistral models. Users should use GPT, Claude, Gemini models, or other models for these tasks.
-
Layered Assistants: The Publish at Path feature is currently under internal testing and should not be used at this time. The AI Generate feature for DESCRIPTION and DECISION GUIDES is not working, and the CometAI team is working internally to resolve the issue.
-
CometAI File Upload - Currently the Sort By functionality causes the File Upload Module to close. The Team is working to rectify this bug.
-
Web Search - The following ASPIRE models are not compatible with the Web Search Functionality: GPT 20B, Mistral 7B, Gemma 4-31B, Llama 3.3 Nemotron 49B.
-
Assistant Workflows and the Visual Workflow Builder are an experimental feature. The internal developer team at OIT is still testing the feature and exploring practical use cases. Behavior, options, and recommended patterns may change. Please proceed with caution and treat workflows as early-access tooling rather than a fully finalized production capability.