Comment by abdullahkhalids
15 hours ago
If you are building your own harness that does correct citations, is the correct thing to give AI access to some deterministic tool that allows it to actually copy paste parts of documents its reading (with links), rather than stochastic reproduction that they do by default?
I did something similar for structured text extraction. I added markers throughout each source document and then, for each piece of info I wanted, I asked the LLM to provide two separate fields:
The latter was just the node number. So then my code could extract the exact snippet, instead of trusting the LLM to quote something verbatim.
That’s what I did when I built my stuff. I have deterministic content with AI commentary, where it seems most people are doing this crazy thing of sending data through the model. I can’t understand it.
You can just look at nouswise or nblm. They do have such harness behind.