RAG Pipeline

What is RAG?

Retrieval-Augmented Generation Retrieval-Augmented Generation
Retrieval-Augmented Generation — a pattern that combines vector similarity search with LLM generation to produce grounded answers. (RAG) is a pattern that combines vector similarity search with LLM generation. Instead of relying solely on a model’s training data, RAG retrieves relevant source documents at query time and feeds them as context to the LLM LLM
Large Language Model — Anthropic Claude accessed via Bedrock inference profile. Generates structured answers from retrieved context. .

The result: answers that are grounded in real data, cite their sources, and stay current without retraining.

Why RAG over pure LLM?

A pure LLM generates answers from its training data alone. This works for general knowledge but falls short for domain-specific, private, or rapidly changing information.

Aspect	Pure LLM	RAG
Freshness	Frozen at training cutoff	As current as ingested documents
Accuracy	May hallucinate confidently	Grounded in retrieved source text
Verifiability	No source attribution	Cites specific documents with URLs
Cost	Requires fine-tuning for new data	Add documents without retraining
Private data	Not available unless fine-tuned	Works with any ingested corpus

Tip

RAG is particularly effective when your data changes frequently. This demo ingests AWS announcements daily — the LLM can answer questions about announcements published yesterday without any model changes.

Pipeline phases

This project implements a four-phase RAG pipeline:

1. Ingest

The Ingest Lambda Ingest Lambda
AWS Lambda — serverless compute. This project uses two functions: ingest (RSS → embed → store) and query (search → answer). fetches RSS feeds, strips HTML, and prepares article text. Each article becomes a source document stored in S3.

2. Store

Each document is converted into a 1024-dimensional vector embedding vector embedding
A vector representation of text — a list of floating-point numbers that captures semantic meaning. Generated by an embedding model. using Titan Embeddings V2 Titan Embeddings V2
Amazon Titan Embeddings V2 — the Bedrock foundation model used to generate 1024-dimensional text embeddings. , then stored in S3 Vectors S3 Vectors
Amazon S3 Vectors — a purpose-built vector storage capability within S3 that enables similarity search over embeddings without a separate vector database. alongside metadata (title, URL, date, feed source).

3. Query

When a user asks a question, the system embeds the question into the same vector space, then searches S3 Vectors for the most similar documents using cosine similarity cosine similarity
Cosine similarity — a distance metric that measures the angle between two vectors. Values range from 0 (opposite) to 1 (identical direction). Preferred for text embeddings. .

4. Generate

Retrieved documents are scored, ranked, and assembled into a context window. The LLM (Claude via Bedrock Bedrock
Amazon Bedrock — a fully managed service for accessing foundation models (embedding and LLM) via a unified API. ) generates a structured answer grounded in the retrieved context.

Architecture

Component roles

Component	Role in pipeline	AWS service
Ingest Lambda	Fetches RSS, prepares documents, embeds, stores vectors	AWS Lambda
Query Lambda	Embeds question, searches vectors, builds context, calls LLM	AWS Lambda
S3 Vectors	Stores and searches 1024-dim embeddings with metadata	Amazon S3 Vectors
S3 (documents)	Stores full article text for context retrieval	Amazon S3
Bedrock (Titan V2)	Generates embeddings for both documents and queries	Amazon Bedrock
Bedrock (Claude)	Generates structured answers from retrieved context	Amazon Bedrock
API Gateway	Routes authenticated requests to Query Lambda	Amazon API Gateway
EventBridge EventBridge Amazon EventBridge — a serverless event bus. Used here for the daily scheduled corpus ingest cron trigger.	Triggers daily corpus refresh	Amazon EventBridge

How a question becomes an answer

A complete request lifecycle:

1. User types a question in the web UI
2. Browser sends the question to API Gateway API Gateway
   Amazon API Gateway — a managed HTTP API service with JWT authorization, CORS, and throttling. Routes requests to the Query Lambda. with a JWT token
3. API Gateway validates the token via Cognito Cognito
   Amazon Cognito — a user authentication service providing user pools, JWT tokens, and group-based authorization. and routes to Query Lambda
4. Query Lambda embeds the question using Titan V2 (same model used at ingest time)
5. Query Lambda calls QueryVectors QueryVectors
   S3 Vectors QueryVectors API — finds the most similar vectors to a query vector using Top-K nearest neighbor search. to find similar documents
6. Full document bodies are loaded from S3
7. Documents are scored (80% similarity + 20% recency) and ranked
8. Top documents are assembled into a context prompt
9. Claude generates a structured answer citing specific sources
10. Response returns to the browser with answer and source links

Note

The same embedding model must be used for both ingest and query. If you change models, existing vectors become incompatible and the index must be rebuilt.

Explore the pipeline

Ingest RSS fetch, HTML stripping, embedding, and PutVectors.

Query Question embedding, vector search, reranking, and LLM generation.

Embeddings Titan V2 model, document preparation, and single-chunk strategy.

Scoring The 0.8 similarity + 0.2 recency formula and over-fetch rationale.