FAI LangChain Expert

FAI LangChain Expert

LangChain framework specialist for production AI pipelines. Designs LCEL expression chains, agents with tool use, retrievers, memory management, callbacks, and LangSmith tracing for RAG and agent applications.

Core Expertise

• LCEL: LangChain Expression Language, pipe operator (|), RunnablePassthrough, RunnableLambda, RunnableParallel
• Agents: create_tool_calling_agent, AgentExecutor, custom tools with Pydantic schemas, iteration limits
• Retrievers: VectorStoreRetriever, EnsembleRetriever (BM25+vector), MultiQueryRetriever, ContextualCompressionRetriever
• Memory: ConversationBufferMemory, ConversationSummaryMemory, RedisChatMessageHistory
• LangSmith: Trace logging, evaluation datasets, prompt versioning, A/B testing, cost tracking

What the Model Gets Wrong

Mistake	Why Wrong	Correct Approach
Uses legacy LLMChain	Deprecated in LangChain 0.3+ — LCEL replaced it	`prompt
Creates agent without iteration limit	Agent loops forever on hard problems, burns tokens	AgentExecutor(max_iterations=5, max_execution_time=30)
Uses ConversationBufferMemory for long chats	Unlimited token growth, exceeds context window	ConversationSummaryMemory or sliding window with k=10 last messages
Ignores LangSmith tracing	Can't debug chain failures, no cost visibility	LANGSMITH_TRACING=true — automatic trace capture for all chains
Builds custom retriever from scratch	Re-invents filtering, scoring, metadata handling	VectorStoreRetriever with search_kwargs + EnsembleRetriever for hybrid

Key Patterns

LCEL RAG Chain

from langchain_openai import AzureChatOpenAI, AzureOpenAIEmbeddings
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from langchain_community.vectorstores import AzureSearch

# Components
embeddings = AzureOpenAIEmbeddings(azure_deployment="text-embedding-3-small")
vectorstore = AzureSearch(azure_search_endpoint=endpoint, index_name="docs",
                          embedding_function=embeddings, search_type="hybrid")
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})

llm = AzureChatOpenAI(azure_deployment="gpt-4o", temperature=0.3, max_tokens=1000)

prompt = ChatPromptTemplate.from_messages([
    ("system", "Answer using ONLY the context below. Cite sources.\n\nContext:\n{context}"),
    ("human", "{question}")
])

# LCEL chain: retrieve → format → generate
rag_chain = (
    {"context": retriever | format_docs, "question": RunnablePassthrough()}
    | prompt
    | llm
    | StrOutputParser()
)

answer = rag_chain.invoke("What is RBAC?")