[BUG] openrouter error

[yimingzhaomath]

Describe the bug
When using instant playlist, it errors out and falls back to genre search. Note that it does not happen every single time. Therefore, it must be related to either my input to generate the playlist, or what the AI returns.

To Reproduce
This is what I put in the box "soft pop songs featuring female vocals"

Expected behavior
The playlist generates without issue.

Screenshots

🎵 NEW MCP-BASED PLAYLIST GENERATION
Request: 'soft pop songs featuring female vocals'
AI Provider: OPENAI
🤖 Using MCP Agentic Workflow for playlist generation
Target: 100 songs
Available tools: song_similarity, artist_similarity, song_alchemy, ai_brainstorm, search_database
============================================================
ITERATION 1/5
Current progress: 0/100 songs
============================================================
--- AI Decision (Iteration 1) ---
Using timeout: 300 seconds for OpenAI/Mistral request
❌ AI tool calling failed: OpenAI error: Expecting ',' delimiter: line 1 column 15 (char 14)
🔄 Fallback: Trying genre search...
   Fallback added 100 songs
============================================================
ITERATION 2/5
Current progress: 100/100 songs
============================================================
✅ Target reached! Stopping iteration.
✅ SUCCESS! Generated playlist with 100 songs
   Total songs collected: 100
   Iterations used: 2/5
   Tools called: 1
📊 Tool Contribution (Collected → Final Playlist):
   • search_database(no filters): 100 songs

Environment (please complete the following information):

• OS: MacOS
• Deployment: OrbStack
• AudioMuse-AI Version: v0.9.0
• Jellyfin/Navidrome Version: Navidrome 0.60.2
• Browser: safari
• CPU: Apple Silicon M4
• RAM: 16GB
• Disk: stock SSD

Additional context

Again, sometimes it works, sometimes it does not. I ran the same thing 2 mins prior to this and it worked.

[NeptuneHub]

I don’t understand: the same query some times work and sometime not? Or different one?
Also which LLM model are you using?
Can you try with llama3.1:8b ?

Anyway if something that only sometimes happen is probably related to the LLM that sometimes go out of the playground.

[yimingzhaomath]

Yes. With the same query, it sometimes works and sometimes not.

I use openrouter/free

It's probably better to use the Gemini free tier than openrouter/free.

[NeptuneHub]

We always try to improve the AI prompt to avoid empty/fallback response, but exist model more or less strongh and query more or less difficult.

Let's start that the AI have this tools avaiable, as you can see in the log you shared:

Available tools: song_similarity, artist_similarity, song_alchemy, ai_brainstorm, search_database

For example a query like this:

Similar song to Artist - Title

should be enough easy because the AI just need to use the song_similarity tools.

Or when you do a query like this:

Danceable happy songs with 120 BPM

should be also be medium because he have the search_database tools but you're searching for multiple value.

All this tools, except ai_brainstorm, require to the AI to practically use an API to get a list or Artist and Title and do and SQL query on a database.

When you search for

soft pop songs featuring female vocals

you don't have an API for that, so AI go to the ai_brainstorm tools that means that the AI need totally to invent the list of Artist and Title. This usually works only with very big model like Gemini and similar.

All of this to say that difficulty of query + power of the model + you song collection impact the results.

For query that ask Moods + Genre + Instrument/vocal, if you have CLAP enabled, I suggest also to try the text search functionality.

So this is not a bug that we can solve but a continuos improvement that we will keep doing.

I can promise nothiong on this, but if you would like to share which query give a good result for your and which a bad result, they could be used for future test.

Thanks.