Sprint Report AI Assistant - Amazon Bedrock RAG

title: "Sprint Report AI Assistant - Amazon Bedrock RAG" date: "2025-10-20" readTime: "8 min" category: "AI" tags: ["Amazon Bedrock", "Python", "Streamlit", "ChromaDB", "RAG", "Agent Architecture", "Embeddings", "Vector Database", "Retrieval", "Chatbot"] excerpt: "Developed Sprint Report AI Assistant using Amazon Bedrock RAG architecture to analyze 11 UN-Habitat QOLI sprint reports, implementing vector embeddings (Titan), semantic search (ChromaDB), and agentic tools - reducing information retrieval time by 80%." image: "/images/blog/screenshot-homepage.png" featured: true

Introduction

At UN-Habitat, managing and extracting insights from 11 comprehensive Quality of Life Impact (QOLI) sprint reports was a time-consuming challenge. I developed the Sprint Report AI Assistant using Amazon Bedrock's RAG (Retrieval Augmented Generation) architecture to transform how the team accesses critical information - reducing information retrieval time by 80%.

🔗 Links:

Sprint Assistant Homepage

Architecture Overview

The system is built on a modern RAG architecture leveraging AWS Bedrock's powerful AI capabilities:

Key Components

Amazon Bedrock Foundation Models
- Claude 3 for natural language understanding and response generation
- Titan Embeddings for converting text into high-dimensional vector representations
- Seamless integration with AWS infrastructure
ChromaDB Vector Database
- Efficient semantic search across all sprint reports
- Fast similarity matching with optimized indexing
- Persistent storage for 11 QOLI sprint reports
Agent Architecture
- Intelligent query routing and processing
- Context-aware response generation
- Agentic tools for enhanced capabilities
Streamlit Frontend
- Clean, intuitive user interface
- Real-time interaction with the AI assistant
- Deployed for easy team access

Implementation Details

Vector Embeddings with Amazon Titan

import boto3

bedrock_runtime = boto3.client('bedrock-runtime')

def generate_embedding(text: str):
    """Generate vector embedding using Amazon Titan"""
    response = bedrock_runtime.invoke_model(
        modelId="amazon.titan-embed-text-v1",
        body=json.dumps({"inputText": text})
    )
    return json.loads(response['body'].read())['embedding']

Semantic Search with ChromaDB

import chromadb
from chromadb.config import Settings

# Initialize ChromaDB client
client = chromadb.Client(Settings(
    persist_directory="./chroma_db"
))

# Create collection for sprint reports
collection = client.create_collection(
    name="sprint_reports",
    metadata={"description": "UN-Habitat QOLI Sprint Reports"}
)

# Add documents with embeddings
collection.add(
    documents=sprint_reports,
    embeddings=embeddings,
    metadatas=metadata_list,
    ids=document_ids
)

RAG Query in Action

Key Features

1. Intelligent Query Understanding

The system leverages Claude 3's advanced natural language understanding to:

Parse complex user queries
Identify relevant context from sprint reports
Generate accurate, contextual responses

2. Semantic Search

ChromaDB enables:

Fast similarity search across thousands of document chunks
Relevant context retrieval in milliseconds
High-precision matching for technical queries

3. Agentic Tools

The agent architecture provides:

Dynamic tool selection based on query type
Multi-step reasoning for complex questions
Fallback mechanisms for edge cases

Sprint Summary Generation

Performance Impact

The Sprint Report AI Assistant delivered measurable improvements:

80% Reduction in information retrieval time
11 Sprint Reports fully indexed and searchable
Sub-second query response times
High Accuracy in extracting relevant information

Technical Stack

Amazon Bedrock - Foundation model hosting and management
Python - Core application development
Streamlit - Interactive web interface
ChromaDB - Vector database for semantic search
RAG Architecture - Retrieval augmented generation pattern
Agent Architecture - Intelligent query processing
Titan Embeddings - High-quality vector representations
Claude 3 - Natural language understanding and generation

Challenges and Solutions

1. Document Chunking Strategy

Challenge: Sprint reports varied significantly in structure and length. Solution: Implemented adaptive chunking based on document sections with overlap to maintain context.

2. Query Precision

Challenge: Generic queries sometimes returned too broad results. Solution: Added query refinement layer and relevance scoring to filter results.

3. Context Window Management

Challenge: Balancing retrieved context with token limits. Solution: Implemented dynamic context selection prioritizing most relevant chunks.

Best Practices Applied

Modular Architecture: Separated embedding, retrieval, and generation layers for maintainability
Persistent Storage: ChromaDB's persistence ensures fast startup and data consistency
Error Handling: Comprehensive error handling for API calls and edge cases
User Feedback: Simple interface for reporting incorrect responses
Monitoring: Logging query patterns to improve system over time

Real-World Impact

The Sprint Report AI Assistant transformed how the UN-Habitat team works:

Instant access to insights across all sprint reports
Faster decision-making with accurate information
Reduced manual search time by 80%
Improved team productivity and collaboration

Future Enhancements

Multi-language support for global teams
Integration with additional UN-Habitat data sources
Advanced analytics on query patterns
Export capabilities for generated insights
Fine-tuning on domain-specific terminology

Conclusion

Building the Sprint Report AI Assistant demonstrates the power of combining Amazon Bedrock's foundation models with ChromaDB's vector search capabilities. The RAG architecture ensures accurate, contextual responses while maintaining sub-second performance.

This project showcases how modern AI can solve real organizational challenges, making information more accessible and teams more productive.

Try it yourself: