Suggestions
Kevin Yang
Co-founder & CEO of include.ai
Highlights
Took some time to reflect on how AI has changed my workflow as the co-founder of an AI startup and wanted to share some concrete ways it's improved my day to day work!
-
I'm now using a combo of Claude and GitHub Co-Pilot to help me write code. This one's been the most surprising for me because I've pretty much stopped using ChatGPT. Especially for frontend code, I can't imagine life without it.
-
What's worked really well for me is passing in a fair amount of example code and instead of having Claude write from scratch, I have it iterate starting from an existing file.
-
Claude's also been super instrumental for data analysis, and other complex tasks that are central to our AI platform's capabilities. It has significantly accelerated our R&D efforts.
-
Notion AI by the Notion team is the best thing I've seen yet in terms of managing notes, outlining drafts, keeping logs, etc. It's the tool I've needed that helps me approach my writing process systematically.
-
Recently adopted Adaline by Arsh Shah Dilbagi who have built the best tool to manage prompts, especially for the complex agent work we've been working on.
Would love to hear the best tools others are using!!
4 advanced tips to improve your AI customer support chatbot!! (Part 2/3)
Today's theme is: Getting Technical with Custom Prompts and Retrieval
1️⃣ LLM based support chatbots work by searching through your external knowledge base using a technique called Retrieval Augmented Generation (RAG) to find the most relevant paragraphs to answer a customer’s question. Chatbots fail when they don’t find anything or they find the wrong content. To maximize the likelihood your chatbot finds the right information, we recommend splitting text (chunking) by markdown headings and also embedding each section with the heading.
2️⃣ Previously we talked about how Question → Answer pairs are the most effective content for chatbots. This tip is also a lot of work, but writing down the question(s) that each block of text answers will greatly increase retrieval accuracy. Think Jeopardy style: “This doc answers how does retrieval augmented generation work.” You can put these questions in the meta data of the document (have your engineer do it).
3️⃣ Collect a good breadth of questions to create a test for your chatbot. Every time you change how your chatbot’s prompt or how you retrieve data, run it against the test to make sure you don’t regress!
4️⃣ Implement real time feedback to see if customers are successful or not. Periodically go through conversations that are not successful and see what content you’re missing.
Like always, wanna chat? Email me at [email protected]!
