docdb-rag-deep-dive

Amazon DocumentDB RAG Deep Dive

An in-depth, guided notebook that builds a complete RAG (Retrieval-Augmented Generation) chatbot using Amazon DocumentDB as the single store for both source documents and vector embeddings. Unlike a typical quickstart, this notebook explains why each parameter matters and demonstrates the impact of changing them with targeted examples.

Most RAG tutorials show you the "happy path." This notebook goes deeper:

• Single-store architecture - vectors and source text live in the same document, eliminating the multi-database sync problem
• HNSW parameter tuning - measures recall vs. latency across different m, efConstruction, and efSearch values so you can see the tradeoffs
• Chunk size and overlap experiments - demonstrates what happens when chunks split mid-sentence and how overlap preserves context
• RAG vs. no-RAG comparison - side-by-side output showing what the LLM gets right (and wrong) without retrieval
• Production resilience patterns - rate limiting, circuit breakers, and retry logic for Bedrock API calls

Check out the AWS Events YouTube channel for a walkthrough of this demo.

Architecture

• Amazon DocumentDB (single collection with HNSW index)
• Amazon Titan Text Embeddings V2
• Anthropic Claude Haiku 4.5 via Amazon Bedrock
• Gradio chat interface
• Amazon SageMaker or any Jupyter-compatible environment

Notebook Sections

#	Section	Purpose
—	Required Configuration	Secret name, AWS region, connection mode
1	Install Dependencies	Required Python packages
2	Import Libraries	Core imports
3	Resilience Utilities	Rate limiter and circuit breaker decorators
4	Connect to DocumentDB	Secrets Manager credentials, HNSW index creation
5	Load and Chunk Documents	PDFs, blog posts, and doc pages → chunked text
6	Generate Embeddings and Insert	Titan embeddings → batch insert into DocumentDB
7	Single-Store vs. Multi-Store	Benchmarks the single-collection approach against a multi-collection pattern
8	HNSW Parameter Tuning	Recall vs. latency across index configurations
9	Configure LLM and Vector Store	Claude Haiku + DocumentDB vector store setup
10	Chunk Overlap - Why It Matters	Same document chunked with/without overlap, then queried
11	Prompt Template	System prompt for grounded Q&A
12	RAG vs No-RAG Comparison	Side-by-side LLM output with and without retrieval
13	Chat Configuration	History length and retrieval limits
14	Launch Chatbot	Gradio chat interface with caching and resilience
15	Cleanup	Close connections

Prerequisites

• AWS account with access to Amazon Bedrock (Titan Text Embeddings V2 and Claude Haiku 4.5)
• Amazon DocumentDB cluster (version 5.0+)
• AWS Secrets Manager secret containing your Amazon DocumentDB credentials
• Python 3.10+

Sample Data

This notebook ingests PDF documents as its knowledge base. You can use any PDFs, but to reproduce the demo as shown, download:

1. Amazon DocumentDB Developer Guide (PDF)
2. Data Modeling with Amazon DocumentDB (PDF)

Place the PDF files in the same directory as the notebook.

Setup

1. Install dependencies:

   pip install -r requirements.txt

2. Download the Amazon DocumentDB TLS certificate:

   wget https://truststore.pki.rds.amazonaws.com/global/global-bundle.pem

3. Download the sample PDFs (see Sample Data above) into this directory.

4. Open docdb-rag-deep-dive.ipynb and update the Required Configuration cell at the top:

   • secret_name — your Secrets Manager secret name
   • aws_region — your AWS region
   • is_bastion — set to 'y' if connecting through a bastion host

5. Run the notebook cells sequentially.