if you don’t or less knowledge about RAG, check it out blog first - link unless you can’t get topic

What is Query Decomposition

Query decomposition is advance RAG techniques used to breaking down complex problem into multiples sub-problems, basically we make less abstract problem. because of multiple sub-problem we get more relevant chunk from retrieval that improve accuracy of llm response.

Decomposition Methods

now we are going to see most common Query Decomposition Techniques are used in RAG.

1. Parallel Decomposition

This methods is used when components of complex query is independent to each other. goal is to maximize retrieval context and efficiency.

Step 1: User Query

• What it is: This is the starting point. A user asks a complex question that may contain multiple sub-topics or require information from various sources to be answered completely. In your diagram, the user query is "What is Machine learning".

Step 2: Query Decomposition

• What it is: The user's query is sent to a large language model (LLM). The LLM's job is to analyze the complex question and break it down into smaller, simpler, and independent sub-questions.

• What the diagram shows: The LLM takes "What is Machine learning" and generates three separate sub-queries: "What is learning", "What is Machine", and "What is Machine learning". This is an example of the LLM trying to get a more comprehensive set of facts by querying different aspects of the original topic.

Step 3: Parallel Retrieval

• What it is: Each of the generated sub-queries is sent simultaneously to a vector database. This is a critical step that makes this method efficient. The system does not wait for one sub-query to be answered before starting the next.

• What the diagram shows: Three separate arrows leave the LLM, each with a different query, and they all go to the "Vector Database." The vector database's role is to perform a similarity search for each query and retrieve the most relevant documents (or "chunks") from its indexed knowledge base. For example:

  • Generated query 1: "What is learning" retrieves documents about the general concept of learning.

  • Generated query 2: "What is Machine" retrieves documents about the concept of a machine.

  • Generated query 3: "What is Machine learning" retrieves core definitions and descriptions of the field itself.

Step 4: Context Aggregation

• What it is: The documents retrieved for each sub-query are collected. Since this is a parallel process, the system now has a broader and more comprehensive set of information than it would have from a single search.

• What the diagram shows: The retrieved documents from all three queries (Query res 1, Query res 2, Query res 3) are combined and sent to the LLM. This gives the LLM a rich, multi-faceted context to work with.

Step 5: Final Response Generation

• What it is: The LLM takes all the retrieved context (from all the sub-queries) and the original user query as its input. Its final task is to read, understand, and synthesize all this information into a single, coherent, and detailed answer.

• What the diagram shows: The LLM processes the combined context and generates a "Final Response" that directly and comprehensively addresses the user's original question, "What is Machine learning."

This complete process ensures that the system doesn't miss any critical information and can provide a more accurate and well-rounded answer than a basic RAG system that would only perform a single search on the original query.

2. Iterative/Multi-hop Query Decomposition

This method is used when components of a complex query require sequential reasoning, where one step is completed based on the previous step's completion.

Step 1: Initial Decomposition (Hop 1)

The goal is to find the first dependent piece of information.

• Input to LLM: The original complex User Query.

• LLM Action: The LLM analyzes the question and generates the first logical sub-question.

• Generated Query 1: "Who invented the light bulb?"

• Retrieval: The system searches the Vector Database using Query 1.

• Output (Retrieved Docs 1): The system finds and extracts the fact: (Thomas Edison).

Step 2: Contextual Refinement (Hop 2)

The system uses the new fact to narrow the search. This is the dependency that defines multi-hop.

• Input to LLM: The User Query is combined with the first retrieved fact (Thomas Edison).

• LLM Action: The LLM uses this combined context to formulate the next logical, specific query.

• Generated Query 2: "What country was Thomas Edison born in?" (Notice how the name from Hop 1 is now in the query).

• Retrieval: The system searches the Vector Database using Query 2.

• Output (Retrieved Docs 2): The system finds and extracts the fact: (United States).

Step 3: Final Targeted Retrieval (Hop 3)

The system uses the results of the second hop to generate the final search query.

• Input to LLM: The User Query is now combined with two retrieved facts (Thomas Edison + United States).

• LLM Action: The LLM formulates the final search query to get the ultimate answer.

• Generated Query 3: "What is the capital of the United States?"

• Retrieval: The system searches the Vector Database using Query 3.

• Output (Retrieved Docs 3): The system finds and extracts the final fact: (Washington, D.C.).

Step 4: Final Synthesis

The LLM integrates all the pieces of the puzzle to generate a comprehensive final answer.

• Input to LLM: The final input box contains the original User Query plus all three facts (Thomas Edison + United States + Washington, D.C.).

• LLM Action: The LLM reads all the context and synthesizes it into a single, cohesive, and logically structured response.

• Final Response: "The inventor of the light bulb was Thomas Edison, who was born in the United States. The capital of that country is Washington, D.C." 💡

Resources

1. langChain academy - GitHub

2. langChain academy - Youtube

---

Learned something? Hit the ❤️ to say “thanks!” and help others discover this article.

Check out my blog for more things related GenAI