Please feel free to checkout the video presentation of this project: YouTube

My goal was to mimic how a human would approach the problem of customer support:

The bot approached this in the following way:

1. Try to answer based on their own knowledge

2. If they can’t, do a web search to fill in the gaps

3. See if it or a similar question had been answered by the community

4. If they're still stumped, escalate

I leveraged the ReAct Agent framework within llama index to create an Orchestrator model that:

- Ingested Documentation & Discord conversations to embed and do a vector search via activeloop

- Rerank the most relevant documents via cohere on the initial query.

This set the foundation of “knowledge” about the problem for the bot which could then:

- Reason through the query + available info and determine what other data it needed

- Leverage it’s tool set (Searching the web, Discord conversations or Documentation) to get more information

For demo purposes, I included:

- A user could email an email address that hit the bot’s endpoint

- The bot went through a multi-step process where it iteratively tried tools to get enough information to answer

- If it answered, a Zendesk ticket would be setup marking the ticket as answered - If it couldn’t answer, a ticket would be setup assigning it to a human.