AI-Powered Tag Recommendation Feature

[chaoliu719]

Feature Overview

I've implemented a comprehensive AI-driven tag recommendation system that seamlessly integrates with the Memos platform. This feature uses Large Language Model (LLM) APIs, such as those from OpenAI, Anthropic, and Gemini, to analyze memo content and suggest relevant tags.

Code & Demo

You can explore my code and see the feature in action via the links below:

• GitHub Branch: https://github.com/chaoliu719/memos/tree/tag-recommand
• Demo Video:

tag.recommand.demo.mov

How to Use

Please jump to https://github.com/orgs/usememos/discussions/5000#discussioncomment-14148770

[chaoliu719]

Wow, it's #5000!

[boojack]

@chaoliu719 I like this approach to introducing AI: starting with a small tag suggestion feature, which is also simple to use. So I think this is a good start, and we can expand the AI capabilities later as needed.

However, for code implementation, we should try to keep the PRs small. We can first introduce llm/ai into the plugin package, then implement the API/service in backend, and finally implement the tag suggestion in frontend.

[chaoliu719]

Thank you for the feedback!

You're right about the architecture - I'll refactor the AI logic from internal/ai/ to
plugin/ai/ to match the established patterns.

Regarding the PR approach, I see your point about smaller commits. Looking at my
commit history, I did bundle the AI client implementation with the API service layer
in one commit, when they could have been separated. I'll structure future AI features
with cleaner separation.

One thing I'm considering: Should we add an AI management section to the Settings
page? Currently, the AI configuration relies on environment variables (AI_BASE_URL,
AI_API_KEY, etc.), but we could allow admins to configure this directly in the UI.

An AI settings section could provide:

• Enable/disable AI features
• Configure AI parameters (base_url, model, api_key)
• Test connectivity
• Set usage limits

This would align with how other plugins are configured and make the feature more
accessible. What's your take on this direction?

[boojack]

Looking into the long term, we’ll definitely need an AI-related setting both in UI and database. In addition to the options(enable/disable, parameters...) you mentioned, it may also require supporting multiple AI models like others do.

However, maybe we can start with just a single environment for now, which would pass a JSON file or string with a standard structure. Once the main functionality(tags suggestion) is ready to use, we can then think about adding a setting to store it into database. For any features, we should start it as simple as possible and improve it later.

[Tyler-Losinski]

I really do love this feature, but would also really like to see an Ollama integration as the standard. With Memos already being a self-hosted option, most likely care about their privacy, and I feel AI would likely also fall into this camp.
Awesome work though, and I hope to see this added one day.

[chaoliu719]

Yes, I totally agree! Supporting Ollama is a great fit, especially since many Memos users care about privacy and self-hosting.

I'm wondering if it's currently possible to properly connect to Ollama just by changing the base_url. Since Ollama might allow an empty key, I'd probably just remove the key validation.
Thanks for the feedback; it's really helpful.

[chaoliu719]

Just completed a major upgrade to the AI tag recommendation system. This release transforms a basic environment variable-based configuration into a comprehensive, production-ready AI integration with full UI support.

tag-recommand-settings-demo.mov

What's new

Previous Implementation:

• Environment variable configuration only (AI_BASE_URL, AI_API_KEY, AI_MODEL)
• Fixed system prompt with no customization
• No validation or error feedback mechanism

Current Implementation:

• Complete workspace settings integration with UI management
• Configurable system prompts and rate limiting
• Connection testing and validation
• Local AI service support (Ollama-friendly)
• Proper error handling and user feedback

Key Features

1. Configuration Management

• UI-driven settings with validation
• Timeout configuration (5-60s with auto-fallback)
• Optional API key for local services

2. Tag Recommendation

• Content analysis with existing tag reuse preference
• Customizable system prompts
• Rate limiting (configurable requests per minute)
• Multi-language support

3. Connection Testing

• Real-time API endpoint validation
• Comprehensive error reporting
• Authentication verification

4. Local AI Support

• API key optional for local services
• Automatic local service detection
• Simplified configuration workflow

[chaoliu719]

@Tyler-Losinski you can test it now, it is very easy to modify the base_url for testing :)

[chaoliu719]

@boojack I refactored the structure of the commit, from the most basic AI capabilities, to the tag recommendation feature, to AI testing, and finally some minor fixes and changes.

Please allow me to explain why I introduced the settings feature now:

1. We can more conveniently test the connectivity of LLM suppliers, or test whether our local converter is working properly, without restarting the entire docker.
2. The request frequency limit for tag recommendations in the settings can help administrators save their tokens.
3. I removed the preset system prompt because I suddenly felt that not everyone is suitable for my own prompt, and everyone should be allowed to customize their own prompt.

[chaoliu719]

Thanks everyone for the feedback. It seems this feature isn't getting much attention at the moment, so I'll be closing this discussion for now.

I'll continue to maintain the AI functionality in my fork. If you're interested, check out the progress here: https://github.com/chaoliu719/memos/tree/tag-recommand