HealthAI
A helpful, grounded health insurance chatbot
The Problem
Call center agents supporting health insurance members face a difficult task:
Answering complex, nuanced
questions about plan benefits in real-time while a member waits on the line.
Project Goals
Allow insurance agents to quickly answer complex health insurance queries without clicking through PDFs or websites.
1
Reduce Call Times
Make it easy to surface information across multiple facets such as copays, deductibles, coinsurance, and coverage limits.
2
Reduce Information Overload
Give agents a "superpower" which allows them to give precise dollar amounts for out of pocket costs and treatment follow-ups.
3
Increase Accuracy
Other Constraints
| Constraint | Brief | Rationale |
|---|---|---|
| Regulatory & Compliance | No member PII in queries | The chatbot must process & transmit plain questions only - no member names, IDs, or conditions should be sent over the wire. |
| Grounding & Truthfulness | Answers must be grounded in source data | The chatbot cannot hallucinate or infer benefits not explicitly documented; responses must cite actual plan attributes |
| Multi-plan Support | Agents handle queries across multiple plan ids | The chatbot should be able to query the correct plan context and retrieve relevant details |
| Follow-up Question Support | Complex queries may require multiple follow-up refinements | The chatbot should be able to handle interruptions and refinements to queries before retrieving the necessary data |
V2 Improvements
| Improvement | Description |
|---|---|
| Cost | FalkorDB and embeddings allow for far less input token usage w/ OpenAI |
| Latency | Related, the smaller payloads dramatically improve latency |
| Accuracy | The LangGraph nodes allow for semantic searches against the plan data as well as retrieving the appropriate sources |
| Retrieval & Caching | V2 narrows down what's relevant, and only sends the context necessary for the query. It also aggressively caches at multiple layers. |
