1️⃣ Hybrid Search

🔹 What is Hybrid Search?

Hybrid search combines lexical (keyword-based) search and semantic (vector-based) search to retrieve documents.

It answers:

“What documents match the words and the meaning of the query?”

🔹 How Hybrid Search Works

1. Keyword search (BM25 / TF-IDF) retrieves exact matches

2. Vector search retrieves semantically similar content

3. Scores are combined or merged

4. Top-N documents are returned

🔹 Why Hybrid Search Exists

• Keyword search → precise but brittle

• Vector search → semantic but approximate

Hybrid search fixes:

• Synonym mismatch

• Domain-specific terminology

• Rare or technical keywords

🔹 Strengths

• Best recall across diverse corpora

• Handles exact terms + meaning

• Industry standard for enterprise RAG

🔹 Weaknesses

• Does not handle redundancy

• Ranking quality still approximate

• Top results may overlap heavily

🔹 Role in RAG

Hybrid search is primarily a candidate generator.

Its job is recall, not precision.

2️⃣ Maximal Marginal Relevance (MMR)

🔹 What is MMR?

MMR (Maximal Marginal Relevance) is a diversity-aware selection algorithm applied after retrieval.

It answers:

“Which documents are relevant without repeating the same idea?”

🔹 How MMR Works

MMR selects documents iteratively by:

• Maximizing relevance to the query

• Minimizing similarity to already selected documents

It balances:

• Relevance

• Diversity

Using a tunable parameter (λ).

🔹 Why MMR Exists

Pure similarity retrieval often returns:

• Near-duplicate chunks

• Rephrased versions of the same paragraph

MMR fixes:

• Redundant context

• Narrow perspective

🔹 Strengths

• Reduces repetition

• Improves coverage

• Very effective for RAG context building

🔹 Weaknesses

• Slight computational overhead

• Does not deeply understand semantics

• Diversity may reduce focus if overused

🔹 Role in RAG

MMR is a context diversification step.

Its job is breadth, not precision ranking.

3️⃣ Reranking

🔹 What is Reranking?

Reranking is a precision optimization step that reorders retrieved documents using a stronger relevance model.

It answers:

“Which documents are actually the best answers to this query?”

🔹 How Reranking Works

1. Take top-N retrieved documents

2. Score each document against the query using:

   • Cross-encoder

   • Transformer

   • LLM

3. Reorder and select top-K

🔹 Why Reranking Exists

Embedding similarity:

• Is geometric

• Misses nuance

• Cannot model deep intent

Reranking fixes:

• Subtle intent mismatch

• Contextual errors

• Weak top-K quality

🔹 Strengths

• Highest precision

• Strong semantic understanding

• Major reduction in hallucination

🔹 Weaknesses

• Slower

• Costly

• Applied only to small candidate sets

🔹 Role in RAG

Reranking is a quality gate before the LLM.

Its job is precision, not recall.

🧠 Core Differences (Conceptual)

🧠 How They Work Together in RAG (Important)

In production-grade RAG, these are not alternatives — they are layers.

Typical Industrial Flow:

1. Hybrid search → broad candidate recall

2. MMR → remove redundancy & expand coverage

3. Reranking → select best final context

4. LLM generation

🎯 When to Use What

Use Hybrid Search when:

• Corpus is large and diverse

• Exact terms matter

• You need strong recall

Use MMR when:

• Retrieved chunks are repetitive

• You want multi-angle answers

• Context window is limited

Use Reranking when:

• Answer quality matters more than latency

• Queries are complex

• Hallucination risk must be minimized

🧠 Interview-Ready One-Liners

• Hybrid search finds candidates

• MMR removes repetition

• Reranking chooses the best answers

🧩 Final Summary

Hybrid search maximizes recall, MMR maximizes diversity, and reranking maximizes precision.
Together, they form the retrieval intelligence layer of a high-quality RAG system.