Conversational AI for bibliometrics requires external structural constraints (Model Context Protocols) to prevent hallucinations and ensure methodological validity.
Interesting though I wonder if APIs are maximum freedom since the design of the api also limits what is practically possible. Eg I ran into this trying to use the OpenAlex api.
I'm also curious how using LLMs to vibe code analysis or dash boards falls into your category. I've been recently doing it with great success...
I'm also wondering about MCP servers but isnt one way to think of them is there are just wrappers around the api? They encode even more restrictions of course by the designer of the mcp server
Ah yes, API definitely have technical limitations! LLMs do mess a lot with my neat diagram, hah.
I think that’s what I mean with Conversational bibliometric. It works well as long as you’re an expert user but you do need less expertise than the API/GBQ (GBQ is SO powerful, and conversational GBQ is even better).
Yes, the MCPs are mostly wrappers, which is why the AI metascientist I’m envisioning would also be supported by deterministic workflows. They would be calling the MCPs, so nobody wastes their time with useless, dangerous, or even methodological invalid queries.
Interesting though I wonder if APIs are maximum freedom since the design of the api also limits what is practically possible. Eg I ran into this trying to use the OpenAlex api.
I'm also curious how using LLMs to vibe code analysis or dash boards falls into your category. I've been recently doing it with great success...
I'm also wondering about MCP servers but isnt one way to think of them is there are just wrappers around the api? They encode even more restrictions of course by the designer of the mcp server
Ah yes, API definitely have technical limitations! LLMs do mess a lot with my neat diagram, hah.
I think that’s what I mean with Conversational bibliometric. It works well as long as you’re an expert user but you do need less expertise than the API/GBQ (GBQ is SO powerful, and conversational GBQ is even better).
Yes, the MCPs are mostly wrappers, which is why the AI metascientist I’m envisioning would also be supported by deterministic workflows. They would be calling the MCPs, so nobody wastes their time with useless, dangerous, or even methodological invalid queries.