TravelMate

Inspiration

We came a long way from Argentina to SF, and most of the times we introduce ourselves people say "hey I always wanted to go to Argentina! What do you guys recommend?"....and we don't know where to start. Challenges for tourists include different exchange rates, lots of tourist traps, overwhelming tourist guides found online...all these make really hard to plan a trip without spending hours and hours doing research.

What it does

TravelMate assist the user in the challenging task of planning a trip by asking questions about preferences as well as loading user data such as flight tickets. After a series of interactions, the agent is able to search online for relevant information and interact with the user in search for converging in the desired result: an itinerary containing a mutually-exclusive-collectively-exhaustive list of activities to make the most of the trip!

How we built it

We wanted to get to a PoC as soon as possible, so we started from the autogpt "vanilla" agent. Then we iterated from there, customizing the agent for our use case. In the process, we integrated the different capabilities & prompts we needed to make the agent become what we needed (loading flight ticket to see flight dates, browsing online for travel recommendations, asking the user for interests, etc).

Challenges we ran into

We found that leaving too many degrees of freedom left too much space for hallucination and the agent losing focus or not being straight to the point. We also had some technical challenges such as rate limits from OpenAI API, parsing problems with the agent's output not containing all necessary keys (plan, criticism or speak were missing some times).

Accomplishments that we're proud of

We are proud of having solved technical issues that were quite discouraging and figured out how to customize the agent to the extend to which we needed.

What we learned

We learned that when building an agent that will serve a specific purpose, we need to to reduce the degrees of freedom by providing concrete commands on things we know are always needed (such as loading flight ticket information). We also learned that starting from a decent working agent and iterating over it is much better than starting from scratch (which was our first approach). Getting up to a point where you have a working PoC as soon as possible is the right way of tackling the challenge of building a custom autonomous agent.

What's next for TravelMate

We will use it for planning our next week in New York! We want to meet more builders and see what the AI community is like in the East Coast. In terms of building, we may iterate more up to an MVP for real users, but we still need to build a UI and improve in terms of UX before so.

Built With

• autogpt
• langchain
• openai