Tech Study Guide

Advanced RAG

Advanced retrieval-augmented generation with hybrid retrieval, query planning, multi-hop retrieval, reranker architectures, GraphRAG, context compression, citation verification, agentic RAG, retrieval evals, and RAG security.

machine-learning rag retrieval advanced

Advanced RAG

Advanced RAG is a retrieval system, ranking system, prompt assembly system, and verification system. The generator is only the last stage.

Command Examples

For one bad answer, capture:
  query
  rewritten query
  top-k chunks
  reranked chunks
  final context
  answer
  citations
  expected source

Example output and meaning:

Command Example output What it does
Captured fields Named fields with concrete values: IDs, scores, tokens, routes, states, timestamps, or errors. Turns a capture template into evidence you can compare across runs.

Advanced Patterns

Pattern Use Risk
Hybrid retrieval Combine lexical and vector search. Score fusion and tuning complexity.
Query planning Break a question into retrieval subqueries. Over-planning and extra latency.
Multi-hop retrieval Retrieve evidence across multiple documents. Missing bridge entities or compounding errors.
Reranker architectures Cross-encoder or LLM reranking. Latency and overfitting to benchmark style.
GraphRAG Use entities and relationships. Graph extraction quality and stale edges.
Context compression Reduce tokens while preserving evidence. Dropping critical qualifiers.
Citation verification Check claims against sources. Requires claim segmentation and source alignment.
Agentic RAG Let an agent decide retrieval actions. Tool loops, cost, and harder evals.

Retrieval Evaluation Dataset

Field Purpose
Query User-facing question.
Relevant source IDs Ground truth for Recall@k.
Required claims What the answer must support.
Forbidden sources ACL and tenant isolation checks.
Freshness version Ensures index recency is tested.
Difficulty slice Exact term, semantic, multi-hop, ambiguous, adversarial.

Practical Lab: Citation Verification

answer_claim: "CloudNativePG performs rolling PostgreSQL minor updates."
cited_source: docs/databases/postgres/cloudnativepg
verifier_question: "Does the cited source support the exact claim?"
labels: supported | partially_supported | unsupported

Study Cards

Question

Why use hybrid retrieval?

Answer

Lexical search catches exact terms and identifiers while vector search catches semantic similarity.

Question

What is multi-hop retrieval?

Answer

Retrieving multiple pieces of evidence that must be connected to answer one question.

Question

Why verify citations?

Answer

A cited document can be related without supporting the exact generated claim.

References