• Rolando@lemmy.world
    link
    fedilink
    English
    arrow-up
    11
    arrow-down
    2
    ·
    4 months ago

    They use rule-based subsystems, but they also use deep neural approaches. The system diagram on their website (in the “How Nibble uses generative AI” video towards the bottom) shows the following:

    The “fact gatherer” and “GPT generative atom” components explicitly use GPT. The “NLU Intent and Entity Models” are classifier components that probably use deep neural approaches, but they don’t say. It’s a good system architecture, it combines the strengths of neural approaches with the strengths of rule-based approaches to address their particular domain, instead just using some massive LLM that tries to do everything.

    Of course whether a website actually NEEDS a negotiation chat bot is another question…

    • Pyro@programming.dev
      link
      fedilink
      English
      arrow-up
      8
      ·
      4 months ago

      That does look like an interesting architecture, if implemented correctly. Based on this observation:

      You can confuse the bot by saying two numbers.

      The NLU component could be a NER model. However, if the following is true:

      The bot will simply offer you its lowest price if you go a few rounds with it.

      It kinda makes the whole system moot in my opinion.