Multi-query retrieval for better recall in LangChain - Step-by-Step Execution

Concept Flow - Multi-query retrieval for better recall

User Input Query

↓

Split into Multiple Queries

↓

Send Each Query to Retriever

↓

Collect Multiple Results

↓

Combine Results

↓

Return Enhanced Recall Output

The process splits one query into several, retrieves results for each, then combines them to improve recall.

Execution Sample

LangChain

queries = ['query part 1', 'query part 2']
results = []
for q in queries:
    res = retriever.get_relevant_documents(q)
    results.extend(res)
combined = combine_results(results)

This code runs multiple queries separately, collects all results, and merges them for better recall.

Execution Table

Step	Action	Query	Retriever Output	Results Collected
1	Split main query	Full query → ['query part 1', 'query part 2']	N/A	[]
2	Send first query	query part 1	Docs A, B	[Docs A, B]
3	Send second query	query part 2	Docs C	[Docs A, B, C]
4	Combine results	N/A	N/A	[Docs A, B, C]
5	Return combined output	N/A	N/A	[Docs A, B, C]

💡 All queries processed and results combined for final output.

Variable Tracker

Variable	Start	After Step 2	After Step 3	Final
queries	[]	['query part 1', 'query part 2']	['query part 1', 'query part 2']	['query part 1', 'query part 2']
results	[]	[Docs A, B]	[Docs A, B, C]	[Docs A, B, C]
combined	N/A	N/A	N/A	[Docs A, B, C]

Key Moments - 3 Insights

Why do we split one query into multiple smaller queries?

Splitting helps the retriever focus on different aspects, increasing the chance to find more relevant documents, as shown in steps 2 and 3 where separate queries return different documents.

How do we avoid duplicate documents in combined results?

After collecting results from each query, we can filter duplicates before combining, ensuring the final list (step 4) has unique documents.

What happens if one query returns no results?

The results list still collects documents from other queries; an empty result for one query doesn't stop the process, as seen in step 3 where results are added only if found.

Visual Quiz - 3 Questions

Look at the execution table, what documents are collected after step 3?

A [Docs A, B]

B [Docs A, B, C]

C [Docs C]

D []

💡 Hint

Check the 'Results Collected' column at step 3 in the execution table.

At which step does the retriever process the second query?

A Step 1

B Step 2

C Step 3

D Step 4

💡 Hint

Look at the 'Action' and 'Query' columns in the execution table for step 3.

If the first query returned no documents, how would the 'results' variable look after step 2?

A []

B [Docs A, B]

C [Docs C]

D ['query part 1']

💡 Hint

Refer to the 'results' variable values in variable_tracker after step 2.

Concept Snapshot

Multi-query retrieval splits one question into parts.
Each part is sent separately to the retriever.
Results from all parts are collected and combined.
This improves recall by covering more angles.
Duplicates should be removed before final output.

Full Transcript

Multi-query retrieval means breaking a single user query into multiple smaller queries. Each smaller query is sent to the retriever to find relevant documents. The results from all queries are collected and combined to improve the chance of finding all useful information. This process helps recall because different queries can catch different relevant documents. After collecting, duplicates are removed to keep the output clean. If one query returns no results, others still contribute to the final combined output.

Practice

1. What is the main purpose of multi-query retrieval in langchain?

easy

A. To store data in multiple vector stores

B. To reduce the number of queries sent to the database

C. To speed up the response time by using a single query

D. To ask several related questions to get better search results

Solution

1. Step 1: Understand the concept of multi-query retrieval

   Multi-query retrieval involves asking multiple related questions instead of just one to improve the quality of search results.

2. Step 2: Identify the main goal of this approach

   The goal is to get better and more complete search results by combining answers from several queries.

3. Final Answer:

   To ask several related questions to get better search results -> Option D

4. Quick Check:

   Multi-query retrieval = multiple related questions for better results [OK]

Hint: Think: multiple questions improve search quality [OK]

Common Mistakes:

• Confusing multi-query with reducing query count
• Thinking it stores data in multiple places
• Assuming it speeds up by using fewer queries

2. Which of the following is the correct way to create a multi-query retriever in langchain?

easy

A. MultiQueryRetriever(vectorstore, llm=llm)

B. MultiQueryRetriever(queries=["query1", "query2"])

C. MultiQueryRetriever.from_llm(vectorstore.as_retriever(), llm)

D. MultiQueryRetriever.from_llm(vectorstore, llm)

Solution

1. Step 1: Check the required parameters for MultiQueryRetriever

   The MultiQueryRetriever is created using from_llm which requires a BaseRetriever (from vectorstore.as_retriever()) and an llm.

2. Step 2: Identify the correct syntax

   MultiQueryRetriever.from_llm(vectorstore.as_retriever(), llm) correctly uses the classmethod with proper arguments.

3. Final Answer:

   MultiQueryRetriever.from_llm(vectorstore.as_retriever(), llm) -> Option C

4. Quick Check:

   Correct syntax includes from_llm, .as_retriever(), and llm [OK]

Hint: Look for from_llm(vectorstore.as_retriever(), llm) syntax [OK]

Common Mistakes:

• Passing queries as a list to constructor
• Using direct constructor without llm_chain
• Forgetting .as_retriever() on vectorstore

3. Given the following code snippet, what will be the output type of results?

retriever = MultiQueryRetriever.from_llm(vectorstore.as_retriever(), llm)
results = retriever.get_relevant_documents("AI")

medium

A. A single document matching the first query only

B. A list of documents combining results from both queries

C. A dictionary with queries as keys and documents as values

D. An error because get_relevant_documents requires no arguments

Solution

1. Step 1: Understand MultiQueryRetriever behavior

   It [uses the LLM to] runs multiple queries and combines their results into a single list of documents.

2. Step 2: Analyze the method call

   Calling get_relevant_documents("AI") triggers retrieval for all [generated] queries, returning combined documents.

3. Final Answer:

   A list of documents combining results from both queries -> Option B

4. Quick Check:

   Multi-query retrieval returns combined document list [OK]

Hint: MultiQueryRetriever returns combined list of documents [OK]

Common Mistakes:

• Expecting a dictionary instead of a list
• Thinking it returns only first query results
• Assuming get_relevant_documents takes no arguments

4. Identify the error in this code snippet using multi-query retrieval:

retriever = MultiQueryRetriever.from_llm(vectorstore, llm)
results = retriever.get_relevant_documents("AI")

medium

A. vectorstore.as_retriever() is missing

B. get_relevant_documents requires a query argument

C. vectorstore is not defined

D. MultiQueryRetriever does not exist in langchain

Solution

1. Step 1: Check the constructor parameters

   from_llm expects a BaseRetriever as the first argument, but vectorstore is a VectorStore; use vectorstore.as_retriever().

2. Step 2: Verify other parts of the code

   get_relevant_documents has argument; llm assumed defined; class exists.

3. Final Answer:

   vectorstore.as_retriever() is missing -> Option A

4. Quick Check:

   from_llm requires BaseRetriever [OK]

Hint: from_llm needs vectorstore.as_retriever() [OK]

Common Mistakes:

• Passing vectorstore directly instead of retriever
• Thinking get_relevant_documents needs no arg when it does
• Assuming vectorstore undefined without context

5. You want to improve recall for a complex topic by using multi-query retrieval with three related queries. Which approach best ensures you get the most relevant documents without duplicates?

hard

A. Use MultiQueryRetriever.from_llm(retriever, llm) to generate multiple queries and deduplicate results

B. Run each query separately and pick the first result from each

C. Combine all queries into one long string and run a single query

D. Use only the query with the highest expected relevance

Solution

1. Step 1: Understand multi-query retrieval goal

   It aims to combine results from multiple related queries to improve recall.

2. Step 2: Identify best practice to avoid duplicates

   Using MultiQueryRetriever.from_llm(retriever, llm) generates multiple queries via LLM and deduplicates results for completeness without repeats.

3. Final Answer:

   Use MultiQueryRetriever.from_llm(retriever, llm) to generate multiple queries and deduplicate results -> Option A

4. Quick Check:

   MultiQueryRetriever generates + deduplicates = best recall [OK]

Hint: MultiQueryRetriever generates queries then deduplicates for best recall [OK]

Common Mistakes:

• Running queries separately without combining results
• Merging queries into one string loses query specificity
• Using only one query misses relevant documents