Similarity search vs MMR retrieval in LangChain

Introduction

Similarity search helps find items close to your query. MMR retrieval balances closeness and variety to avoid repeats.

When you want to find documents or data points most like your question.

When you want diverse results that cover different aspects without repeating the same info.

When building chatbots that should give varied answers, not just the closest match.

When searching a large database and you want both relevance and variety.

When you want to avoid showing very similar items multiple times.

Syntax

LangChain

from langchain.vectorstores import FAISS

# Similarity search
results = vectorstore.similarity_search(query, k=5)

# MMR retrieval
results = vectorstore.max_marginal_relevance_search(query, k=5, fetch_k=10)

similarity_search returns the top k closest matches to the query.

max_marginal_relevance_search returns k results balancing similarity and diversity, using fetch_k candidates internally.

Examples

Finds the 3 most similar documents to 'What is AI?'.

LangChain

results = vectorstore.similarity_search('What is AI?', k=3)

Finds 3 documents that are similar but also diverse, choosing from top 6 candidates.

LangChain

results = vectorstore.max_marginal_relevance_search('What is AI?', k=3, fetch_k=6)

If no close matches exist, similarity search returns the closest available, even if not very similar.

LangChain

results = vectorstore.similarity_search('Nonexistent topic', k=3)

MMR tries to pick diverse results even if all are not very close to the query.

LangChain

results = vectorstore.max_marginal_relevance_search('Nonexistent topic', k=3, fetch_k=5)

Sample Program

This program creates a small vector store from example sentences about AI. It runs both similarity search and MMR retrieval for the query 'Tell me about AI'. It prints the top 3 results from each method so you can see the difference.

LangChain

from langchain.vectorstores import FAISS
from langchain.embeddings import OpenAIEmbeddings

# Sample documents
documents = [
    'AI is the simulation of human intelligence.',
    'Machine learning is a subset of AI.',
    'Deep learning uses neural networks.',
    'AI can be used in healthcare.',
    'Neural networks mimic the brain.'
]

# Create embeddings
embeddings = OpenAIEmbeddings()

# Build vector store
vectorstore = FAISS.from_texts(documents, embeddings)

query = 'Tell me about AI'

# Similarity search
similar_results = vectorstore.similarity_search(query, k=3)

# MMR retrieval
mmr_results = vectorstore.max_marginal_relevance_search(query, k=3, fetch_k=5)

print('Similarity Search Results:')
for doc in similar_results:
    print('-', doc.page_content)

print('\nMMR Retrieval Results:')
for doc in mmr_results:
    print('-', doc.page_content)

Important Notes

Similarity search is fast and finds the closest matches but can return very similar or repeated results.

MMR retrieval adds diversity by balancing similarity and novelty, which is useful when you want varied answers.

MMR is slightly slower because it fetches more candidates internally before picking diverse results.

Summary

Similarity search finds the closest matches to your query.

MMR retrieval balances closeness and diversity to avoid repeated or very similar results.

Use similarity search for quick, relevant results; use MMR when you want variety in answers.

Practice

1. What is the main difference between Similarity search and MMR retrieval in Langchain?

easy

A. Similarity search uses machine learning; MMR retrieval uses rule-based filtering.

B. Similarity search finds closest matches; MMR retrieval balances closeness and diversity.

C. Similarity search always returns random results; MMR retrieval returns exact matches.

D. Similarity search is slower than MMR retrieval in all cases.

Solution

1. Step 1: Understand Similarity Search

   Similarity search finds the closest matches to the query based on similarity scores.

2. Step 2: Understand MMR Retrieval

   MMR retrieval balances between closeness and diversity to avoid repeated or very similar results.

3. Final Answer:

   Similarity search finds closest matches; MMR retrieval balances closeness and diversity. -> Option B

4. Quick Check:

   Difference = Closest vs Balanced [OK]

Hint: Similarity = closeness; MMR = closeness plus diversity [OK]

Common Mistakes:

• Thinking MMR returns only closest matches
• Confusing MMR with random retrieval
• Assuming similarity search ensures diverse results

2. Which of the following is the correct way to initialize an MMR retriever in Langchain?

easy

A. mmr_retriever = MMRRetriever(documents, k=5, lambda_mult=0.7)

B. mmr_retriever = MMRRetriever(documents, top_k=5, lambda=0.7)

C. mmr_retriever = MMRRetriever(documents, k=5, lambda=0.7)

D. mmr_retriever = MMRRetriever(documents, top_k=5, lambda_mult=0.7)

Solution

1. Step 1: Check parameter names for MMRRetriever

   The correct parameters include k for number of results and lambda_mult for diversity weight.

2. Step 2: Identify correct parameter usage

   Options B, C, and D use top_k or lambda which are incorrect parameter names.

3. Final Answer:

   mmr_retriever = MMRRetriever(documents, k=5, lambda_mult=0.7) -> Option A

4. Quick Check:

   Correct params = k and lambda_mult [OK]

Hint: MMR uses k and lambda_mult parameters [OK]

Common Mistakes:

• Using 'top_k' instead of 'k'
• Using 'lambda' instead of 'lambda_mult'
• Mixing parameter names from similarity search

3. Given the following code snippet using Langchain's similarity search and MMR retrieval, what will be the main difference in their outputs?

docs = ["apple pie recipe", "apple tart recipe", "banana bread recipe", "cherry pie recipe"]
query = "apple dessert"
sim_results = similarity_search(docs, query, k=2)
mmr_results = mmr_retrieval(docs, query, k=2, lambda_mult=0.5)

medium

A. sim_results will return the two most similar apple recipes; mmr_results will return one apple and one different dessert recipe.

B. sim_results will return random recipes; mmr_results will return the two closest apple recipes.

C. Both sim_results and mmr_results will return the same two apple recipes.

D. sim_results will return only one recipe; mmr_results will return three recipes.

Solution

1. Step 1: Analyze similarity_search output

   Similarity search returns the top 2 closest matches, likely both apple recipes due to query "apple dessert".

2. Step 2: Analyze mmr_retrieval output

   MMR balances similarity and diversity, so it will likely return one apple recipe and one different dessert (like cherry pie) to increase variety.

3. Final Answer:

   sim_results will return the two most similar apple recipes; mmr_results will return one apple and one different dessert recipe. -> Option A

4. Quick Check:

   Similarity = closest; MMR = closest + diverse [OK]

Hint: Similarity = closest matches; MMR adds variety [OK]

Common Mistakes:

• Assuming both return identical results
• Thinking MMR returns random results
• Confusing number of results returned

4. You wrote this code to perform MMR retrieval but get an error:

mmr = MMRRetriever(documents, k=3, lambda_mult=0.6)
results = mmr.get_relevant_documents(query)

What is the likely cause of the error?

medium

A. Calling 'get_relevant_documents' instead of 'retrieve_documents'.

B. Using 'k' instead of 'top_k' when initializing MMRRetriever.

C. Missing import statement for MMRRetriever.

D. Using 'lambda_mult' instead of 'lambda' parameter.

Solution

1. Step 1: Check parameter names in MMRRetriever initialization

   The parameters k and lambda_mult are correct.

2. Step 2: Confirm method name correctness

   get_relevant_documents is a valid method for retrievers.

3. Final Answer:

   Missing import statement for MMRRetriever. -> Option C

4. Quick Check:

   Missing import = NameError [OK]

Hint: Import MMRRetriever before use [OK]

Common Mistakes:

• Forgetting import statement for MMRRetriever
• Confusing method names
• Wrong parameter names like top_k or lambda

5. You want to retrieve 3 documents related to "climate change" using Langchain. You want results that are relevant but also cover different aspects like causes, effects, and solutions. Which retrieval method and parameters should you choose?

hard

A. Use MMR retrieval with k=3 and lambda_mult=0 to get only closest matches.

B. Use similarity search with k=3 for closest matches only.

C. Use similarity search with k=1 repeated 3 times to get variety.

D. Use MMR retrieval with k=3 and lambda_mult=0.8 to balance relevance and diversity.

Solution

1. Step 1: Identify need for relevance and diversity

   You want documents that are relevant but cover different aspects, so diversity is important.

2. Step 2: Choose retrieval method and parameters

   MMR retrieval balances relevance and diversity using lambda_mult. A value like 0.8 prioritizes relevance while introducing some diversity.

3. Final Answer:

   Use MMR retrieval with k=3 and lambda_mult=0.8 to balance relevance and diversity. -> Option D

4. Quick Check:

   Diversity + relevance = MMR with 0 < lambda_mult < 1 [OK]

Hint: Lower lambda_mult increases diversity in MMR [OK]

Common Mistakes:

• Using similarity search only for diverse results
• Setting lambda_mult to 0 disables relevance
• Repeating similarity search calls for variety