Innovating Search Experience with Amazon OpenSearch and Amazon Bedrock

Source: https://bigdataboutique.com/blog/innovating-search-experience-with-amazon-opensearch-and-amazon-bedrock-d045bc
Author: Big Data Boutique

Summary

A practical guide to building production semantic search and RAG using Amazon OpenSearch Service for vector storage and Amazon Bedrock (Claude, Titan) for embeddings and generation.

Architecture Overview

User Query
    ↓
Amazon Bedrock (Titan Embeddings) → query_vector
    ↓
Amazon OpenSearch Service (kNN index) → top-k chunks
    ↓
Amazon Bedrock (Claude) → generated answer
    ↓
User Response

This is a fully AWS-native RAG stack.

OpenSearch Vector Index Setup

import boto3
from opensearchpy import OpenSearch
 
# Create index with vector field
index_body = {
    "settings": {
        "index.knn": True
    },
    "mappings": {
        "properties": {
            "embedding": {
                "type": "knn_vector",
                "dimension": 1536,
                "method": {
                    "name": "hnsw",
                    "space_type": "l2",
                    "engine": "nmslib"
                }
            },
            "content": {"type": "text"},
            "source": {"type": "keyword"}
        }
    }
}

Bedrock Titan Embeddings

bedrock = boto3.client("bedrock-runtime")
 
def embed(text):
    response = bedrock.invoke_model(
        modelId="amazon.titan-embed-text-v1",
        body=json.dumps({"inputText": text})
    )
    return json.loads(response["body"].read())["embedding"]

Hybrid Search in OpenSearch

OpenSearch supports BM25 + kNN hybrid:

{
  "query": {
    "hybrid": {
      "queries": [
        {"match": {"content": "query text"}},
        {"knn": {"embedding": {"vector": [...], "k": 10}}}
      ]
    }
  },
  "search_pipeline": {
    "phase_results_processors": [{
      "normalization-processor": {
        "combination": {"technique": "rrf"}
      }
    }]
  }
}

RAG Generation with Claude

def rag_answer(query, retrieved_chunks):
    context = "\n\n".join([c["content"] for c in retrieved_chunks])
    
    response = bedrock.invoke_model(
        modelId="anthropic.claude-3-sonnet-20240229-v1:0",
        body=json.dumps({
            "messages": [{
                "role": "user",
                "content": f"Context:\n{context}\n\nQuestion: {query}"
            }]
        })
    )
    return json.loads(response["body"].read())["content"][0]["text"]

Production Considerations

1. Chunking strategy: split documents before embedding (see Text Chunking)
2. Embedding cost: Titan at $0.0001/1K tokens — budget for large corpora
3. Index warm-up: OpenSearch kNN needs warm-up after restart
4. Metadata filtering: OpenSearch supports pre-filter on non-vector fields

Related Articles

• Migrating to Elasticsearch with Dense Vector for Carousell Spotlight — similar migration case study
• Chunking Strategies for LLM Applications — chunking for RAG
• Nearest Neighbor Indexes for Similarity Search — HNSW infrastructure

Related Concepts

• RAG — primary pattern
• Dense Vector Retrieval — storage and retrieval
• Semantic Search — enabled capability
• Hybrid Search — BM25 + kNN
• Text Chunking — document preparation
• Agentic Search — potential extension with Bedrock agents