How to Build Custom LLM Chain with Prompt Templates

What this does

Custom LLM chains let you define reusable prompt templates with dynamic variables, few-shot examples, and structured outputs — giving full control over how the LLM is instructed.

Steps

• Define a prompt template with multiple variables. Use ChatPromptTemplate for structured messages.

from langchain.prompts import ChatPromptTemplate
from langchain_ollama import ChatOllama
from langchain.schema.output_parser import StrOutputParser

llm = ChatOllama(model="llama3.2", temperature=0)

prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a {role} expert. Answer concisely."),
    ("human", "{question}")
])

• Create and invoke the chain.

chain = prompt | llm | StrOutputParser()

result = chain.invoke({
    "role": "Python",
    "question": "What is a decorator?"
})
print(result)

• Add few-shot examples. Include example pairs in the prompt.

from langchain.prompts import FewShotChatMessagePromptTemplate

examples = [
    {"input": "What is a list?", "output": "An ordered collection."},
    {"input": "What is a dict?", "output": "A key-value mapping."},
]

example_prompt = ChatPromptTemplate.from_messages([
    ("human", "{input}"),
    ("ai", "{output}"),
])

few_shot_prompt = FewShotChatMessagePromptTemplate(
    example_prompt=example_prompt,
    examples=examples,
)

final_prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a Python tutor."),
    few_shot_prompt,
    ("human", "{question}"),
])

chain = final_prompt | llm | StrOutputParser()
print(chain.invoke({"question": "What is a tuple?"}))

• Use partial variables. Pre-bind variables that don't change per call.

prompt = ChatPromptTemplate.from_template("""
Context: {context}
Question: {question}
Answer:""").partial(context="Default knowledge base content.")

chain = prompt | llm | StrOutputParser()
print(chain.invoke({"question": "What is RAG?"}))

Verification

python -c "
from langchain.prompts import ChatPromptTemplate
pt = ChatPromptTemplate.from_template('Hello {name}')
r = pt.format(name='World')
print(r)
# Expected: Hello World
"

Common failures

• Message role order matters. LangChain validates message order (system, then alternating human/ai). Wrong order raises ValidationError.
• Few-shot example key mismatch. Example dict keys must match example_prompt template variables exactly.
• Partial overwrite ignored. Calling invoke with a key that was already partial-bound silently uses the partial value. Verify by inspecting prompt.input_variables.
• Version mismatch - The installed package or runtime differs from the command shown; check the version first and rerun the smallest verification command.
• Local environment drift - Another service, virtual environment, model, or path is being used; print the active binary path and configuration before changing the guide steps.

Related guides

• How to Create Basic LangChain Chains with LCEL
• How to Compose Multiple Chains with Pipe Operator