Project — Advanced RAG on AWS

TeleNova Support Assistant

A production-grade Retrieval-Augmented Generation pipeline for a fictional telecom's customer support. Ask a question and watch every stage run live — query transformation, hybrid retrieval (dense vectors + BM25), reciprocal-rank fusion, two re-rankers side by side, and a grounded, cited answer with a faithfulness check.

S3 Vectors Hybrid Search RRF Fusion Re-ranking Bedrock RAG

Ask TeleNova support

Pipeline mode

Full pipeline — query transformation, hybrid retrieval, and both re-rankers. Highest quality, ~3.8s.

Enter a question to run the pipeline.

Query Transformation

Claude rewrites the question into retrieval-friendly variants (multi-query · HyDE · step-back)

—

Hybrid Retrieval

Dense ANN over S3 Vectors (semantic) run alongside BM25 (exact keywords)

Dense · S3 Vectors · cosine

—

Sparse · BM25 · keywords

—

Reciprocal Rank Fusion

Merge both ranked lists by rank — items found by both retrievers rise to the top

—

Re-ranking — two engines compared

An LLM-as-judge (in-region) vs a managed cross-encoder (Cohere Rerank, cross-region)

LLM-as-judge · Claude

—

Cross-encoder · Cohere Rerank 3.5

—

Grounded Answer

Claude answers only from the top passages, with inline citations and a faithfulness check

Architecture. Dense vectors live in Amazon S3 Vectors (1024-dim Titan v2 embeddings, cosine) in us-east-1; sparse retrieval is an in-Lambda BM25 over the same chunks. The managed re-ranker (Cohere Rerank 3.5) is called cross-region in us-west-2 because Bedrock Rerank isn't offered in us-east-1 — shown side by side with an in-region LLM-as-judge re-ranker. When the live backend is unavailable (or rate-limited), this page runs a client-side simulation so the pipeline is always explorable.