LangChain / RAG / Vector DBs · Lesson

Extending Retrieval Chains with Custom Logic

Build custom retrieval chains that integrate complex business logic, pre-processing steps, or specialized filtering.

Extending Retrieval Chains with Custom Logic is a free LangChain / RAG / Vector DBs lesson on CoddyKit — lesson 3 of 4. You can read the complete lesson below for free — then practise it hands-on in the browser with a built-in code editor and a 24/7 AI tutor. It is part of the LangChain / RAG / Vector DBs learning path, one of 4 lessons in the course, and your progress syncs across the web and the CoddyKit app.

Beyond Basic RAG

Welcome! In this lesson, we'll dive into extending LangChain's retrieval chains. While standard Retrieval Augmented Generation (RAG) is powerful, real-world applications often need more nuanced control.

We'll learn how to inject custom logic into the retrieval process to make your RAG systems smarter and more tailored to specific needs.

Practical Customization Needs

Why would you need custom logic in a retrieval chain? Consider these common scenarios:

Filtering by User Permissions: Only retrieve documents accessible to the current user.
Prioritizing Fresh Data: Boost documents created or updated recently.
Removing Irrelevant Sections: Clean up retrieved text before passing it to the LLM.
Dynamic Query Rephrasing: Automatically improve user queries for better search results.

These needs go beyond what a basic retriever offers.

LangChain's Custom Primitives

LangChain provides flexible primitives to insert custom Python logic directly into your chains:

RunnableLambda: This allows you to wrap any Python function, making it a runnable component. It's perfect for applying arbitrary transformations.
RunnablePassthrough: This simply passes its input through to the next step. It's useful for injecting new keys into the input dictionary or for identity operations.

These are your building blocks for custom steps.

Enhancing User Queries (Pre-processing)

One powerful customization is query pre-processing. This means modifying the user's input query before it's sent to the retriever or vector store.

You could add specific keywords based on detected intent.
Expand common abbreviations or synonyms.
Rephrase the query to improve embedding search results.

This subtle step can significantly boost the relevance of retrieved documents.

Custom Query Transformer Code

Let's see how to implement a simple query pre-processor using RunnableLambda. This example adds a 'detailed search for' prefix to the original query.

from langchain_core.runnables import RunnableLambda
from langchain_core.promnpts import ChatPromptTemplate
from langchain_openai import ChatOpenAI

# Mock LLM for demonstration purposes
class MockLLM(ChatOpenAI):
    def invoke(self, input):
        return f"LLM processed: {input['enhanced_query']}"

def enhance_query(input_dict):
    original_query = input_dict["question"]
    return {"enhanced_query": f"detailed search for {original_query}"}

llm = MockLLM() # In a real app, use ChatOpenAI(model="gpt-4")

prompt = ChatPromptTemplate.from_template(
    "Answer based on the following search query: {enhanced_query}"
)

custom_chain = (
    {"enhanced_query": RunnableLambda(enhance_query)} # Our custom step
    | prompt
    | llm
)

result = custom_chain.invoke({"question": "latest AI trends"})
print(result)

Refining Retrieved Documents (Post-processing)

Another crucial area for custom logic is document post-processing. This occurs after documents have been retrieved but before they are passed to the LLM.

Filtering: Remove documents that don't meet certain criteria (e.g., outdated, wrong source).
Re-ranking: Reorder documents based on custom relevance scores.
Summarizing: Condense lengthy documents to fit context windows.

This ensures the LLM receives the most relevant and concise context, improving answer quality and reducing token usage.

Filtering Documents by Metadata

Here's an example of filtering retrieved documents based on their metadata. We'll simulate a retriever that returns documents and then filter them to only include those from a specific 'blog' source.

from langchain_core.documents import Document
from langchain_core.runnables import RunnableLambda
from langchain_core.promnpts import ChatPromptTemplate
from langchain_openai import ChatOpenAI

# Mock LLM for demonstration
class MockLLM(ChatOpenAI):
    def invoke(self, input):
        context = input.get("context", "No context provided")
        return f"LLM processed docs: {context}"

# Mock retriever returning Documents with metadata
def mock_retrieve(query):
    return [
        Document(page_content="Doc A about AI", metadata={"source": "blog"}),
        Document(page_content="Doc B about ML", metadata={"source": "research"}),
        Document(page_content="Doc C about AI ethics", metadata={"source": "blog"}),
    ]

def filter_by_source(docs, desired_source="blog"):
    # Only keep documents from the 'blog' source
    return [doc for doc in docs if doc.metadata.get("source") == desired_source]

llm = MockLLM()
prompt = ChatPromptTemplate.from_template(
    "Answer based on the following context: {context}"
)

# Build a simple chain with retrieval and custom filter
custom_retrieval_chain = (
    RunnableLambda(mock_retrieve) # Simulate retrieval
    | RunnableLambda(filter_by_source) # Apply custom filter
    | (lambda docs: {"context": "\n\n".join([d.page_content for d in docs])}) # Format for LLM
    | prompt
    | llm
)

result = custom_retrieval_chain.invoke("AI topics")
print(result)

End-to-End Custom Chain

You can combine both query pre-processing and document post-processing within a single LangChain chain. The flow would look something like this:

User Query
Custom Query Pre-processor
Retriever (e.g., Vector Store)
Custom Document Post-processor
LLM for Answer Generation

This modular approach gives you fine-grained control over every step of your RAG pipeline, making it highly adaptable to complex requirements.

Advanced: Conditional Routing

For even more dynamic behavior, LangChain offers RunnableBranch. This powerful construct allows your chain to take different paths based on certain conditions.

For example, you could:

Use one retriever if the query is about 'code' and another for 'general knowledge'.
Apply different document filters based on the user's role.

RunnableBranch enables sophisticated, context-aware RAG workflows.

Check Your Understanding

Which LangChain primitive is best suited for inserting a simple Python function to modify data (e.g., filter a list of documents) within a chain?

Recap: Extending Retrieval

Great job! In this lesson, you learned how to extend LangChain retrieval chains with custom logic:

We explored the need for customization in real-world RAG.
You discovered RunnableLambda and RunnablePassthrough as key tools.
We saw how to pre-process queries for better retrieval.
You learned to post-process retrieved documents for refined context.
We touched on advanced concepts like RunnableBranch for conditional logic.

Experiment with these techniques to build highly customized and efficient RAG applications!

Frequently asked questions

Is the “Extending Retrieval Chains with Custom Logic” lesson free?

Yes — the full text of “Extending Retrieval Chains with Custom Logic” is free to read here on the web, and the LangChain / RAG / Vector DBs course includes 4 lessons in total. To practise it interactively (a built-in code editor and a 24/7 AI tutor) and unlock the rest of the LangChain / RAG / Vector DBs course, upgrade to CoddyKit PRO.

What will I learn in “Extending Retrieval Chains with Custom Logic”?

Build custom retrieval chains that integrate complex business logic, pre-processing steps, or specialized filtering. You practise LangChain / RAG / Vector DBs with hands-on code you run directly in the browser, and a 24/7 AI tutor answers your questions as you work through the lesson.

Do I need any experience to start LangChain / RAG / Vector DBs?

No prior experience is required. LangChain / RAG / Vector DBs on CoddyKit is structured for beginners through advanced learners; this is — lesson 3 of 4, so you can start here or from the beginning and move at your own pace.

How long does the “Extending Retrieval Chains with Custom Logic” lesson take?

Most CoddyKit lessons take about 5–10 minutes. Each one is bite-sized and interactive, so you make steady progress and pick up exactly where you left off across the web and the app.

Can I write and run code in this LangChain / RAG / Vector DBs lesson?

Yes. Every LangChain / RAG / Vector DBs lesson includes a built-in code editor, so you write and run real code right in your browser and get instant AI feedback — no local setup required.

All lessons in this course

Developing Custom Document Loaders
Integrating Custom Embedding Models
Extending Retrieval Chains with Custom Logic
Building Custom Output Parsers

← Back to LangChain / RAG / Vector DBs