PolicyMind AI: Intelligent Insurance Document Assistant using Gemma 4

CHAVA BRAHMAM — Sun, 10 May 2026 12:14:04 +0000

This is a submission for the Gemma 4 Challenge: Build with Gemma 4

What I Built

I built PolicyMind AI, a Retrieval-Augmented Generation (RAG) based insurance assistant powered by Gemma 4, Gemini Embeddings, and ChromaDB.

Insurance policy documents are often long, complex, and difficult for users to understand. Finding information such as waiting periods, exclusions, surgery coverage, claim conditions, or policy benefits usually requires manually searching through multiple pages of PDFs.

PolicyMind AI solves this problem by allowing users to ask natural language questions directly about insurance documents and receive grounded AI-generated answers with relevant policy citations.

The system:

Extracts text from insurance policy PDFs
Splits documents into semantic chunks
Generates embeddings using Gemini Embeddings
Stores vector embeddings in ChromaDB
Retrieves relevant policy chunks using semantic similarity
Uses Gemma 4 to generate context-aware responses

Instead of returning entire PDFs, PolicyMind AI provides concise, user-friendly answers backed by retrieved document context.

This project demonstrates how Gemma 4 can be used to build domain-specific AI assistants for real-world document intelligence applications.

System Architecture

The above architecture illustrates the Retrieval-Augmented Generation (RAG) workflow used in PolicyMind AI.

Insurance policy documents are embedded and stored inside a vector database.
User queries are semantically embedded and matched against relevant document chunks.
Retrieved policy chunks are passed to Gemma 4 along with the user query to generate grounded responses with citations.

Demo

PolicyMind AI was tested on 7 real insurance policy PDF documents containing information about:

Waiting periods
Surgery coverage
Policy exclusions
Claim conditions
Health benefits
Insurance terms and conditions

The system allows users to ask natural language questions and retrieves grounded answers directly from the policy documents using semantic search and Gemma 4.

Example Question

what is the policy on eye issues?

Sample Output

The generated responses include relevant policy citations, helping users quickly locate the source information inside lengthy insurance documents.

Code

GitHub Repository:
https://github.com/BrahmamChava/PolicyMind-AI

How I Used Gemma 4

PolicyMind AI uses Gemma 4 as the response generation model in the Retrieval-Augmented Generation (RAG) pipeline.

After retrieving the most relevant insurance policy chunks from ChromaDB using semantic similarity search, Gemma 4 generates grounded, citation-aware, and context-aware answers based on the retrieved insurance document content.

For this project, I used:

models/gemma-4-26b-a4b-it
models/gemma-4-31b-it

These models were selected because they provide:

Strong long-context understanding
High-quality instruction-following capabilities
Better reasoning over lengthy insurance documents
Effective grounded response generation for RAG workflows
Improved handling of complex legal and structured document text

Insurance policy documents often contain:

Long legal text
Structured tables
Complex exclusions and conditions
Domain-specific terminology
Multi-page contextual dependencies

Gemma 4 was a strong fit for this use case because its long-context capabilities helped process retrieved insurance chunks more effectively and generate concise user-friendly answers.

The large context handling capability of Gemma 4 is particularly useful in Retrieval-Augmented Generation systems where multiple retrieved chunks from lengthy policy documents must be processed together before generating the final response.

The project combines:

Gemma 4 for grounded response generation
Gemini Embeddings for semantic retrieval
ChromaDB for vector storage
Recursive Character Chunking for improved retrieval precision

This architecture enables users to ask natural language questions and receive grounded answers with relevant policy citations instead of manually reading lengthy insurance PDFs.

Additionally, using Gemma 4 helped improve:

Retrieval grounding quality
Context preservation across retrieved chunks
Response coherence for insurance-related queries
Citation-aware answer generation

DEV Community: CHAVA BRAHMAM