Designing for Retrieval

In the previous lesson, we learned the "Why" of RAG. In this lesson, we master the "How." An Amazon Bedrock Knowledge Base is the managed implementation of the RAG pattern. It automates the extraction, chunking, embedding, and storage of your data.

However, as a Professional Developer, you can't just click "Next" on the default settings. You must design the architecture to match the shape of your data.

---

1. The Anatomy of a Knowledge Base

An Amazon Bedrock Knowledge Base consists of four integrated components:

1. The Data Source: Usually an S3 bucket (though web and database connectors are emerging).
2. The Embedding Model: (e.g., Titan Embeddings v2).
3. The Vector Store: (e.g., Amazon OpenSearch Serverless).
4. The Chunker: The logic that cuts your data into pieces.

---

2. Mastery of Chunking Strategies

This is one of the most technical areas of the AIP-C01 exam. If your "chunks" are too small, they lose context. If they are too large, the AI gets confused and costs rise.

Strategy A: Fixed-Size Chunking

• What it is: Dividing text into blocks of exactly $X$ tokens with an overlap of $Y$ tokens.
• When to use: Short documents, simple FAQs, or when you are on a tight budget.
• The "Overlap": Essential to prevent a sentence from being cut in half. Usually 10-20%.

Strategy B: Hierarchical Chunking (Parent-Child)

• What it is: You store small "child" chunks for retrieval, but when a match is found, you send the larger "parent" chunk to the model.
• When to use: Complex technical manuals where a specific "Step" needs the context of the "Chapter."

Strategy C: Semantic Chunking

• What it is: The system uses a model to detect when a Topic changes and cuts the chunk there.
• When to use: Highly inconsistent documents where some sections are short and others are 20 pages long.

graph TD
    DOC[Document: 10,000 Words] --> CH{Chunking Strategy}
    CH -->|Fixed| F[Chunk 1: 512 tokens]
    CH -->|Fixed| F2[Chunk 2: 512 tokens]
    CH -->|Hierarchical| H[Parent: Chapter]
    H --> C1[Child: Paragraph]
    H --> C2[Child: Table]

---

3. The Ingestion Lifecycle

When you trigger a "Sync" on a Knowledge Base, the following events occur (and can be monitored via CloudWatch):

1. Extraction: Bedrock reads the file from S3.
2. Chunking: The text is split according to your rules.
3. Embedding: Each chunk is converted into a vector.
4. Storage: The vector and its metadata are "Upserted" into the Vector Store.

Professional Developer Tip: Metadata Injection

You can significantly improve RAG performance by including a .metadata.json file alongside your S3 objects. This allows you to "Score" or "Filter" results later.

---

4. Advanced: Managing Custom Parsers

By default, Bedrock uses a standard text extractor. If your documents contain high-density tables or complex charts, you should use Amazon Textract as a custom parser within your Knowledge Base.

Scenario: You have 10,000 medical lab reports. A standard parser will mangle the table of results. Action: Enable the "Advanced Parsing" option in the Knowledge Base settings to use Textract's table detection.

---

5. Decision Matrix: Choosing the Vector Store

Requirement	Preferred Vector Store
Minimum Overhead	Amazon OpenSearch Serverless
Highest Performance Scale	Amazon OpenSearch (Managed Clusters)
Already Use SQL/Postgres	Amazon Aurora (pgvector)
Need Third-Party/Multi-Cloud	Pinecone / MongoDB Atlas

---

6. Real-World Code: Listing Knowledge Bases

import boto3

client = boto3.client('bedrock-agent')

def list_my_kbs():
    response = client.list_knowledge_bases(maxResults=10)
    for kb in response['knowledgeBaseSummaries']:
        print(f"Name: {kb['name']} | ID: {kb['knowledgeBaseId']} | Status: {kb['status']}")

# Triggering a Daily Sync Job
def trigger_sync(kb_id, ds_id):
    client.start_ingestion_job(
        knowledgeBaseId=kb_id,
        dataSourceId=ds_id
    )
    print("Sync started successfully.")

---

Knowledge Check: Test Your Indexing Knowledge

---

Summary

Designing the Knowledge Base is the technical "hard part" of Domain 1. Get the chunking wrong, and your AI will be "stupid" even if the model is "smart." In the next lesson, we move to the final piece of the RAG puzzle: Context Assembly and Semantic Retrieval.

---

Next Lesson: The Precision Search: Context Assembly and Semantic Retrieval