Inspiration
Trying to sift through all the technical information about different plans and eligibility information is complicated and always changing. And even after finding a plan it's difficult to know how good the supported providers are. This agent can query multiple sources of data and interact in natural language to assist the user to filter down from available plans to the ones matching their specific needs.
What it does
Filters available plans by region and then uses a vector search to lookup benefits that the user needs. Afterwards it uses a genie agent to query across mimilabs data to show information on the selected plans.
How we built it
We started a notebook in Databricks to arrange the LangGraph Agent and then published that to a serving endpoint. We created a vector index to be able to lookup plans based on semantic similarity to described/needed benefits and then setup a Genie space to handle querying the plan information across multiple tables.
Challenges we ran into
We weren't familiar with the data set to know where how to find the relevant data across the minilabs dataset. It was our first time setting up a langgraph agent and hosting it in Databricks. We often would stop getting updates working from the notebook until we detached and reattached to compute.
Accomplishments that we're proud of
That we create an orchestrated multi-agent system that was able to use a vector index, genie search and foundation models to gather the needed input from the user and preserve the correct state throughout the conversation.
What we learned
How to model a multi-agent conversation in LangGraph. How to permission access to genie and vector searches in database from a hosted agent.
What's next for Medicare Plan Explorer
Pairing with someone who better understands the medicare plan data will help us tune each agent to be able to retrieve the best plans. In addition to data available in mimilabs we would like to integrate data from google maps or other user reviews to augment suggested providers based on user ratings.
Log in or sign up for Devpost to join the conversation.