Using LangChain Expression Language (LCEL) for prompts and retrieval

In my previous post Use case for RAG and LLM my sample code only used basic string manipulation of the prompt. On this post I will show how to use LangChain Expression Language (LCEL)

Instead of string manipulation, LCEL offers a more effective alternative. Here are the step by step conversion:

• Instead of using python string interpolation:

prompt = f"I need help on {context}"

use the same string without interpolation and a chat prompt template

prompt = ChatPromptTemplate.from_template("I need help on {context}")

• We can directly use the vector store as a retriever within a sub-chain, simplifying the search and integration process.

retriever = vector_store.as_retriever(search_type='similarity')
context_subchain = itemgetter('user_query') | retriever

• Finally combine the prompts, retriever and output processing in a chain. RunnablePassthrough is used for the user_query is supplied when the chain is invoked. itemgetter is use for llm_personality which will be substituted from a disctionary passed on the chain's invocation.

chain = (
    {
        'context': context_subchain, 
        'user_query': RunnablePassthrough(), 
        'llm_personality': itemgetter('llm_personality')
    } 
    | prompt
    | model
    | StrOutputParser()
)

Here is as sample code is written using LCEL

template_system = """

Use the following information to answer the user's query:

{context}
"""

template_user = """
User's query:

{user_query}
"""

prompt = ChatPromptTemplate.from_messages([
        SystemMessagePromptTemplate.from_template(template_system),
        HumanMessagePromptTemplate.from_template(template_user)
    ])

retriever = vector_store.as_retriever(search_type='similarity')
context_subchain = itemgetter('user_query') | retriever

chain = (
    {
        'context': context_subchain, 
        'user_query': RunnablePassthrough(), 
        'llm_personality': itemgetter('llm_personality')
    } 
    | prompt
    | model
    | StrOutputParser()
)

response = chain.invoke({**{'user_query': user_query}, **prompt_placeholders})

You can see a more complete commit diff from old string manipulation to LCEL