implement resource embedding

[soyuka]

suggestion for #652

Basically creates a tool for the llm to retrieve resources:

Also interesting: https://github.com/openai/codex/pull/5239/files

[matteo8p]

@soyuka This is an awesome approach! I would love to have a pre-installed tool to get resources. I have a couple of thoughts on your approach:

• I like how you made additions to MCPClientManager. This is the foundation of MCPJam. I like the new helpers you added. I do think we can change the function names. We already have listResource for single server. Maybe we can replace that function to allow multiple serverIds passed in, or rename your function to listResourcesForMultipleServers to be more explicit
• I don't think we need a resources embedding model. The model requires the user to provide either a openai, deepseek, or gemini API key otherwise this feature wouldn't work. Instead of an embedding model, I think we can just pass in the URI's of the resources in the system prompt. The model should be able to call getResource with the right URI with the URI's in its system prompt.
• In addition, we provide free models via cloud too, so we'll have to make some modifications there too, but we'll take care of that.

This also makes me think whether or not we should add other default tools like web search too, to simulate real LLM environments like Claude Desktop.

[soyuka]

you don't add empty lines usually right? my ide does that I should probably rolbacm

[soyuka]

this is merely a prototype for now I still want to improve the usage but at least it allows me to test resources

[soyuka]

I don't think we need a resources embedding model. The model requires the user to provide either a openai, deepseek, or gemini API key otherwise this feature wouldn't work. Instead of an embedding model, I think we can just pass in the URI's of the resources in the system prompt. The model should be able to call getResource with the right URI with the URI's in its system prompt.

Agreed that this adds quite some complexity, but I felt it was the correct way of handling resources ideally. What about having both implementation, if the configured model doesn't support I can fallback on just giving the resources to the system prompt?

I'm also wondering how I could cache the resource calls as for now it is calling the tool though the client already has the whole resource. Another thing that bothers me in the modelcontextprotocol didn't quite figure the best way of paginating resources (modelcontextprotocol/modelcontextprotocol#799) I may have some ideas but for now I'll leave that aside.