RAG App: Amazon Bedrock + Aurora PostgreSQL

How it works

RAG = Retrieval-Augmented Generation. Instead of relying on the model's training data (which can be wrong or stale), you store your own documents as vectors, retrieve the relevant ones at query time, and pass them to Claude as context.

Your question │ ▼ Bedrock Titan Embeddings V2 → 1024-dim vector │ ▼ Aurora PostgreSQL + pgvector → cosine similarity search → top 3 docs │ ▼ Claude on Bedrock (reads docs as context) │ ▼ Grounded answer

Three AWS services. No hallucination. Fully auditable — every answer traces back to a document you ingested.

Setup guide

Create an Aurora PostgreSQL cluster

Go to RDS → Create database and choose:

Engine: Aurora (PostgreSQL-compatible)
Version: 15.4 or later (pgvector is pre-installed)
Template: Dev/Test
Instance class: db.t3.medium
Initial database name: ragdb

Or with the AWS CLI:

aws rds create-db-cluster \
  --db-cluster-identifier rag-cluster \
  --engine aurora-postgresql \
  --engine-version 15.4 \
  --master-username postgres \
  --master-user-password YOUR_PASSWORD \
  --database-name ragdb \
  --region us-east-1

aws rds create-db-instance \
  --db-instance-identifier rag-cluster-instance \
  --db-cluster-identifier rag-cluster \
  --engine aurora-postgresql \
  --db-instance-class db.t3.medium \
  --region us-east-1

Once the cluster is available, connect and run setup.sql:

psql -h YOUR_CLUSTER_ENDPOINT -U postgres -d ragdb -f setup.sql

-- setup.sql
CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE IF NOT EXISTS docs (
    id        SERIAL PRIMARY KEY,
    content   TEXT    NOT NULL,
    embedding vector(1024)
);

CREATE INDEX IF NOT EXISTS docs_emb_idx
ON docs USING ivfflat (embedding vector_cosine_ops)
WITH (lists = 50);

pgvector is pre-installed on Aurora PostgreSQL 15.2+. No manual extension install needed.

Enable Bedrock model access

Go to Amazon Bedrock → Model access
Click Manage model access
Enable Titan Embeddings V2 and Claude Sonnet
Click Save changes — access is instant

Make sure you're in the same region as your Aurora cluster.

IAM permissions

Your execution environment needs bedrock:InvokeModel on the two models:

{
  "Version": "2012-10-17",
  "Statement": [{
    "Effect": "Allow",
    "Action": ["bedrock:InvokeModel"],
    "Resource": [
      "arn:aws:bedrock:*::foundation-model/amazon.titan-embed-text-v2:0",
      "arn:aws:bedrock:*::foundation-model/anthropic.claude-sonnet-4-6"
    ]
  }]
}

For local dev, run aws configure with a user that has this policy.

Install dependencies

pip install -r requirements.txt

boto3>=1.34.0
psycopg2-binary>=2.9.9

Set environment variables

export DB_HOST=your-cluster.cluster-xxxx.us-east-1.rds.amazonaws.com
export DB_NAME=ragdb
export DB_USER=postgres
export DB_PASSWORD=your-password
export AWS_REGION=us-east-1

Find your cluster endpoint: RDS → Clusters → your cluster → Endpoints → Writer endpoint

Run it

python rag_app.py

============================================================
  RAG App: Amazon Bedrock + Aurora PostgreSQL (pgvector)
============================================================

[1/3] Setting up Aurora pgvector...
Aurora pgvector ready.

[2/3] Ingesting Oracle migration docs...
Ingested 5 documents into Aurora pgvector.

[3/3] Asking: What is the difference between Oracle sequences
      and Aurora PostgreSQL sequences?

------------------------------------------------------------
Answer:
------------------------------------------------------------
In Oracle, you call sequences using .NEXTVAL FROM DUAL.
In Aurora PostgreSQL, use SELECT nextval('seq_name').
No more DUAL table. You can also use SERIAL or
GENERATED AS IDENTITY for auto-increment columns.
------------------------------------------------------------

The code

Three functions. That's the entire RAG pipeline.

embed() — convert text to a vector

def embed(text: str) -> List[float]:
    resp = bedrock.invoke_model(
        modelId="amazon.titan-embed-text-v2:0",
        body=json.dumps({"inputText": text, "dimensions": 1024}),
    )
    return json.loads(resp["body"].read())["embedding"]

ingest() — embed and store documents

def ingest(documents: List[str]):
    with get_db() as conn:
        with conn.cursor() as cur:
            for doc in documents:
                vec = embed(doc)
                cur.execute(
                    "INSERT INTO docs (content, embedding) VALUES (%s, %s::vector)",
                    (doc, vec)
                )
        conn.commit()

ask() — retrieve context, generate answer

def ask(question: str) -> str:
    # 1. Cosine similarity search in Aurora
    q_vec = embed(question)
    with get_db() as conn:
        with conn.cursor() as cur:
            cur.execute("""
                SELECT content FROM docs
                ORDER BY embedding <=> %s::vector
                LIMIT 3
            """, (q_vec,))
            chunks = [row[0] for row in cur.fetchall()]

    # 2. Pass retrieved docs as context to Claude
    context = "\n\n---\n\n".join(chunks)
    resp = bedrock.invoke_model(
        modelId="anthropic.claude-sonnet-4-6",
        body=json.dumps({
            "anthropic_version": "bedrock-2023-05-31",
            "max_tokens": 512,
            "messages": [{"role": "user", "content":
                f"Answer using ONLY this context:\n\n{context}\n\nQuestion: {question}"
            }]
        })
    )
    return json.loads(resp["body"].read())["content"][0]["text"]

Customising for your own docs

Replace ORACLE_MIGRATION_DOCS with your content:

MY_DOCS = [
    "Your first document...",
    "Your second document...",
]
ingest(MY_DOCS)

To load from files:

import os

docs = []
for fname in os.listdir("./my-docs"):
    with open(f"./my-docs/{fname}") as f:
        docs.append(f.read())

ingest(docs)

For documents longer than ~500 words, split them into smaller chunks before ingesting. Each chunk should cover one topic so similarity search stays accurate.

Cost estimate (dev/demo)

Service	Usage	Approx cost
Aurora PostgreSQL db.t3.medium	8 hrs/day	~$1.50/day
Bedrock Titan Embeddings V2	5 docs + 10 queries	< $0.01
Bedrock Claude Sonnet	10 queries	< $0.05

Stop the cluster when not in use to avoid charges: RDS Console → select cluster → Actions → Stop temporarily

Common errors

Error	Fix
could not connect to server	Check DB_HOST env var and security group — inbound port 5432 must allow your IP
AccessDeniedException	Enable model access in Bedrock console (Step 2 above)
could not open extension control file "vector"	Upgrade Aurora to PostgreSQL 15.2+
could not load library "vector"	Same — pgvector requires Aurora PostgreSQL 15.2+

RAG App: Amazon Bedrock+ Aurora PostgreSQL

How it works

Setup guide

The code

embed() — convert text to a vector

ingest() — embed and store documents

ask() — retrieve context, generate answer

Customising for your own docs

Cost estimate (dev/demo)

Common errors

RAG App: Amazon Bedrock
+ Aurora PostgreSQL