Inspiration
As devs the most frustrating thing is not being able to get a timely answer when you have a question. We're hoping to solve that with Agents that approach support queries like a human.
What it does
It leverages an Orchestrator Agent model to take in a query, reason through possible paths to a solution and leverage tools such as querying the Docs, finding related questions in Discord or searching the Web to answer the problem.
We provide it with a "baseline" knowledge of the project / problem statement by doing an initial query on the docs to prime the LLM. If the answer happens to be in there, we'll return the answer.
All queries are then created into a Zendesk ticket, either marking as solved or escalating to a human.
How we built it
- Index with Llama Index
- Store in Vector DB
- Query VDB
- Built Tools to ingest data, clean and query it (Discord, Web, and Docs)
Challenges we ran into
The biggest is that Agents arn't as competent as we'd hoped. Pro: we keep our jobs for now. Con: You have to significantly constrain and provide guardrails to make it work.
Accomplishments that we're proud of
Shipping!
What we learned
The tools and orchestration needs significant guard rails via prompt engineering and conditionals in code to make it useful.
What's next for Autonomous Agent Support
We'll see...
Built With
- deeplake
- llamaindex
- openai
- python
Log in or sign up for Devpost to join the conversation.