Exploring Configurable Parameters

This lesson explains configurable fields — a mechanism for passing runtime parameters to runnables (for example, switching the LLM model used for a specific invocation). Configurable fields let you avoid hard-coding options when initializing a runnable and instead override them per invocation. This is useful for cost control (default to a cheaper model) and flexibility (upgrade to a stronger model when needed).

Why use configurable fields?

Avoid reinitializing runnables to change runtime behavior.
Control costs by selecting cheaper defaults and overriding on demand.
Compose overrides with dynamic prompts and memory for flexible workflows.

Quick overview

Import required classes.
Register a configurable field on a runnable (e.g., model_name).
Compose the runnable with a prompt.
Invoke normally (uses default) or override with with_config(configurable={...}).

Imports and setup

from langchain_core.prompts import PromptTemplate
from langchain_openai import ChatOpenAI
from langchain_core.runnables import ConfigurableField

Register a configurable field on a runnable

Initialize the ChatOpenAI runnable with a default model (for example gpt-3.5-turbo) and expose model_name as a configurable field that can be overridden at invocation time:

model = ChatOpenAI(model_name="gpt-3.5-turbo").configurable_fields(
    model_name=ConfigurableField(
        id="model_name",
        name="model name",
        description="The GPT model to use for chat",
    )
)

Create a prompt template

Wrap the template text in backticks in prose to avoid MDX parsing issues:

prompt = PromptTemplate.from_template("Write a Haiku on {subject}")

Compose the chain and invoke (default)

Compose the prompt and the runnable, then invoke normally. This uses the runnable’s default model (gpt-3.5-turbo):

chain = prompt | model

# Invoke with the default (gpt-3.5-turbo)
result_default = chain.invoke({"subject": "cat"})
print(result_default)

Example output (using the default model):

AIMessage(content="Whiskers soft and fine\nPurring gently in the sun\nGraceful feline friend", response_metadata={'model_name': 'gpt-3.5-turbo', 'token_usage': {'completion_tokens': 19, 'prompt_tokens': 13, 'total_tokens': 32}})

Override the configurable field at runtime

Use with_config(configurable={...}) to pass overrides for registered configurable fields. Keys in the dictionary must match the configurable field ids you registered (in this example, "model_name"):

# Force the chain to use GPT-4 for this invocation
result_gpt4 = chain.with_config(configurable={"model_name": "gpt-4"}).invoke({"subject": "cat"})
print(result_gpt4)

Example output (overridden to use GPT-4):

AIMessage(content="Soft purr in the night,\nEyes gleaming in moon's soft light,\nCat in calm delight.", response_metadata={'model_name': 'gpt-4', 'token_usage': {'completion_tokens': 22, 'prompt_tokens': 13, 'total_tokens': 35}})

The configurable field id you register must match the parameter name the runnable expects at runtime (for example, model_name above). Use with_config(configurable={...}) to override values per invocation.

Practical considerations

Topic	Guidance	Example
Cost management	Default to a cheaper model and override only when needed	Default: `gpt-3.5-turbo`; override to `gpt-4` for complex prompts
Key matching	Ensure configurable field ids match runnable parameter names	Use `model_name` to override the `ChatOpenAI` model parameter
Composition	Combine configurable overrides with dynamic prompt inputs and memory	Use runtime `model_name` + user-specific prompts + stored memory
Debugging	Inspect `response_metadata` to confirm which model was used	Check `response_metadata['model_name']` in the AIMessage

When to use configurable fields

You want per-invocation control over LLM models or other runtime settings.
Your application serves varied tasks that require different capability levels.
You want to avoid reinitializing runnables or changing deployment configuration mid-run.

Links and references

OpenAI Models Overview: https://platform.openai.com/docs/models/overview
LangChain (core concepts and runnables): https://langchain.readthedocs.io/ (or the relevant repo/docs for your distribution)

Try different OpenAI model names from the models list to see how outputs and token usage change. This pattern provides runtime flexibility without changing deployed code.

Introduction

Overview of LangChain

Key Components of LangChain

Interacting with LLMs

Tips, Tricks, and Resources

Introduction to LCEL

Adding Memory to LLM Apps

Performing Retrieval

Implementing Chains

Using Tools

Building Agents

Conclusion

Building Blocks of LLM Apps

Exploring Configurable Parameters

Why use configurable fields?

Quick overview

Imports and setup

Register a configurable field on a runnable

Create a prompt template

Compose the chain and invoke (default)

Override the configurable field at runtime

Practical considerations

When to use configurable fields

Links and references

Watch Video

​Why use configurable fields?

​Quick overview

​Imports and setup

​Register a configurable field on a runnable

​Create a prompt template

​Compose the chain and invoke (default)

​Override the configurable field at runtime

​Practical considerations

​When to use configurable fields

​Links and references

Watch Video

Why use configurable fields?

Quick overview

Imports and setup

Register a configurable field on a runnable

Create a prompt template

Compose the chain and invoke (default)

Override the configurable field at runtime

Practical considerations

When to use configurable fields

Links and references