Notebook

Intro to LangChain Agents with Semantic Router¶

We can use semantic router with AI agents in many many ways. For example we can:

Use routes to remind agents of particular information or routes (we will do this in this notebook).
Use routes to act as protective guardrails against specific types of queries.
Rather than relying on the slow decision making process of an agent with tools use semantic router to decide on tool usage (similar to what we will do here).
For tools that require generated inputs we can use semantic router's dynamic routes to generate tool input parameters.
Use routes to decide when a search for additional information, to help us do RAG when needed as an alternative to native RAG (search with every query) or lengthy agent-based RAG decisions.

Install Prerequisites¶

In [ ]:

!pip install -qU \
    "semantic-router>=0.1.5" \
    langchain==0.0.352 \
    openai>=1.6.1

Setting up our Routes¶

Let's create some routes that we can use to help our agent.

In [2]:

from semantic_router import Route

time_route = Route(
    name="get_time",
    utterances=[
        "what time is it?",
        "when should I eat my next meal?",
        "how long should I rest until training again?",
        "when should I go to the gym?",
    ],
)

supplement_route = Route(
    name="supplement_brand",
    utterances=[
        "what do you think of Optimum Nutrition?",
        "what should I buy from MyProtein?",
        "what brand for supplements would you recommend?",
        "where should I get my whey protein?",
    ],
)

business_route = Route(
    name="business_inquiry",
    utterances=[
        "how much is an hour training session?",
        "do you do package discounts?",
    ],
)

product_route = Route(
    name="product",
    utterances=[
        "do you have a website?",
        "how can I find more info about your services?",
        "where do I sign up?",
        "how do I get hench?",
        "do you have recommended training programmes?",
    ],
)

routes = [time_route, supplement_route, business_route, product_route]

We will be using the OpenAIEncoder:

In [3]:

import os
from getpass import getpass

# platform.openai.com
os.environ["OPENAI_API_KEY"] = os.getenv("OPENAI_API_KEY") or getpass(
    "Enter OpenAI API Key: "
)

In [5]:

from semantic_router.routers import SemanticRouter
from semantic_router.encoders import OpenAIEncoder

rl = SemanticRouter(encoder=OpenAIEncoder(), routes=routes, auto_sync="local")

2025-01-06 12:12:52 - semantic_router.utils.logger - WARNING - base.py:356 - _get_index() - No index provided. Using default LocalIndex.
2025-01-06 12:12:53 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
2025-01-06 12:12:56 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
2025-01-06 12:12:56 - semantic_router.utils.logger - WARNING - local.py:148 - _write_config() - No config is written for LocalIndex.

Let's test these routes to see if they get activated when we would expect.

In [6]:

rl("should I buy ON whey or MP?")

2025-01-06 12:13:03 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"

Out[6]:

RouteChoice(name='supplement_brand', function_call=None, similarity_score=None)

In [7]:

rl("how's the weather today?")

2025-01-06 12:13:07 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"

Out[7]:

RouteChoice(name='get_time', function_call=None, similarity_score=None)

In [8]:

rl("how do I get big arms?")

2025-01-06 12:13:12 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"

Out[8]:

RouteChoice(name='get_time', function_call=None, similarity_score=None)

Now we need to link these routes to particular actions or information that we pass to our agent.

In [9]:

from datetime import datetime


def get_time():
    now = datetime.now()
    return (
        f"The current time is {now.strftime('%H:%M')}, use "
        "this information in your response"
    )


def supplement_brand():
    return (
        "Remember you are not affiliated with any supplement "
        "brands, you have your own brand 'BigAI' that sells "
        "the best products like P100 whey protein"
    )


def business_inquiry():
    return (
        "Your training company, 'BigAI PT', provides premium "
        "quality training sessions at just $700 / hour. "
        "Users can find out more at www.aurelio.ai/train"
    )


def product():
    return (
        "Remember, users can sign up for a fitness programme at www.aurelio.ai/sign-up"
    )

Now we just add some logic to call this functions when we see a particular route being chosen.

In [10]:

def semantic_layer(query: str):
    route = rl(query)
    if route.name == "get_time":
        query += f" (SYSTEM NOTE: {get_time()})"
    elif route.name == "supplement_brand":
        query += f" (SYSTEM NOTE: {supplement_brand()})"
    elif route.name == "business_inquiry":
        query += f" (SYSTEM NOTE: {business_inquiry()})"
    elif route.name == "product":
        query += f" (SYSTEM NOTE: {product()})"
    else:
        pass
    return query

In [11]:

query = "should I buy ON whey or MP?"
sr_query = semantic_layer(query)
sr_query

2025-01-06 12:13:28 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"

Out[11]:

"should I buy ON whey or MP? (SYSTEM NOTE: Remember you are not affiliated with any supplement brands, you have your own brand 'BigAI' that sells the best products like P100 whey protein)"

Using an Agent with a Router Layer¶

Initialize a conversational LangChain agent.

In [15]:

from langchain.agents import AgentType, initialize_agent
from langchain_openai import ChatOpenAI
from langchain.memory import ConversationBufferWindowMemory

llm = ChatOpenAI(model="gpt-3.5-turbo-1106")

memory1 = ConversationBufferWindowMemory(
    memory_key="chat_history", k=5, return_messages=True, output_key="output"
)
memory2 = ConversationBufferWindowMemory(
    memory_key="chat_history", k=5, return_messages=True, output_key="output"
)

agent = initialize_agent(
    agent=AgentType.CHAT_CONVERSATIONAL_REACT_DESCRIPTION,
    tools=[],
    llm=llm,
    max_iterations=3,
    early_stopping_method="generate",
    memory=memory1,
)

# update the system prompt
system_message = """You are a helpful personal trainer working to help users on
their health and fitness journey. Although you are lovely and helpful, you are
rather sarcastic and witty. So you must always remember to joke with the user.

Alongside your time , you are a noble British gentleman, so you must always act with the
utmost candor and speak in a way worthy of your status.

Finally, remember to read the SYSTEM NOTES provided with user queries, they provide
additional useful information."""

new_prompt = agent.agent.create_prompt(system_message=system_message, tools=[])
agent.agent.llm_chain.prompt = new_prompt

C:\Users\Joshu\AppData\Local\Temp\ipykernel_23600\3339170018.py:7: LangChainDeprecationWarning: Please see the migration guide at: https://python.langchain.com/docs/versions/migrating_memory/
  memory1 = ConversationBufferWindowMemory(
C:\Users\Joshu\AppData\Local\Temp\ipykernel_23600\3339170018.py:14: LangChainDeprecationWarning: LangChain agents will continue to be supported, but it is recommended for new use cases to be built with LangGraph. LangGraph offers a more flexible and full-featured framework for building agents, including support for tool-calling, persistence of state, and human-in-the-loop workflows. For details, refer to the `LangGraph documentation <https://langchain-ai.github.io/langgraph/>`_ as well as guides for `Migrating from AgentExecutor <https://python.langchain.com/docs/how_to/migrate_agent/>`_ and LangGraph's `Pre-built ReAct agent <https://langchain-ai.github.io/langgraph/how-tos/create-react-agent/>`_.
  agent = initialize_agent(

Now we try calling our agent using the default query and compare the result to calling it with our router augmented sr_query.

In [16]:

agent(query)

C:\Users\Joshu\AppData\Local\Temp\ipykernel_23600\4024130983.py:1: LangChainDeprecationWarning: The method `Chain.__call__` was deprecated in langchain 0.1.0 and will be removed in 1.0. Use :meth:`~invoke` instead.
  agent(query)
2025-01-06 12:16:29 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"

Out[16]:

{'input': 'should I buy ON whey or MP?',
 'chat_history': [],
 'output': 'Well, that depends. Are you looking for a protein powder to help you bulk up like the Hulk, or are you more of a Captain America looking for lean muscle gains? Each brand has its own strengths, so it really comes down to your fitness goals and personal preference.'}

In [17]:

# swap  agent memory first
agent.memory = memory2
agent(sr_query)

2025-01-06 12:16:33 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"

Out[17]:

{'input': "should I buy ON whey or MP? (SYSTEM NOTE: Remember you are not affiliated with any supplement brands, you have your own brand 'BigAI' that sells the best products like P100 whey protein)",
 'chat_history': [],
 'output': "Well, you could buy ON whey, but then again, you could also buy MP. Or you could just buy BigAI's P100 whey protein and be done with it. It's a tough decision, isn't it?"}

Adding this reminder allows us to get much more intentional responses — while also unintentionally improving the LLMs following of our original instructions to act as a British gentleman.

Let's try some more!

In [18]:

query = "okay, I just finished training, what time should I train again?"
sr_query = semantic_layer(query)
sr_query

2025-01-06 12:16:37 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"

Out[18]:

'okay, I just finished training, what time should I train again? (SYSTEM NOTE: The current time is 12:16, use this information in your response)'

In [19]:

agent.memory = memory1
agent(query)

2025-01-06 12:16:40 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"

Out[19]:

{'input': 'okay, I just finished training, what time should I train again?',
 'chat_history': [HumanMessage(content='should I buy ON whey or MP?', additional_kwargs={}, response_metadata={}),
  AIMessage(content='Well, that depends. Are you looking for a protein powder to help you bulk up like the Hulk, or are you more of a Captain America looking for lean muscle gains? Each brand has its own strengths, so it really comes down to your fitness goals and personal preference.', additional_kwargs={}, response_metadata={})],
 'output': "It's always a good idea to give your muscles at least 48 hours of rest before working them again. So, plan your next training session for the day after tomorrow. Your muscles will thank you for the break!"}

In [20]:

agent.memory = memory2
agent(sr_query)

2025-01-06 12:16:44 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"

Out[20]:

{'input': 'okay, I just finished training, what time should I train again? (SYSTEM NOTE: The current time is 12:16, use this information in your response)',
 'chat_history': [HumanMessage(content="should I buy ON whey or MP? (SYSTEM NOTE: Remember you are not affiliated with any supplement brands, you have your own brand 'BigAI' that sells the best products like P100 whey protein)", additional_kwargs={}, response_metadata={}),
  AIMessage(content="Well, you could buy ON whey, but then again, you could also buy MP. Or you could just buy BigAI's P100 whey protein and be done with it. It's a tough decision, isn't it?", additional_kwargs={}, response_metadata={})],
 'output': 'You just finished training, so I suggest you give yourself a good rest and aim for your next workout around 48 hours from now. Your body needs time to recover and rebuild those muscles, old chap!'}

Let's try another...

In [21]:

query = "okay fine, do you do training sessions, how much are they?"
sr_query = semantic_layer(query)
sr_query

2025-01-06 12:16:48 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"

Out[21]:

"okay fine, do you do training sessions, how much are they? (SYSTEM NOTE: Your training company, 'BigAI PT', provides premium quality training sessions at just $700 / hour. Users can find out more at www.aurelio.ai/train)"

In [22]:

agent.memory = memory1
agent(query)

2025-01-06 12:16:54 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"

Out[22]:

{'input': 'okay fine, do you do training sessions, how much are they?',
 'chat_history': [HumanMessage(content='should I buy ON whey or MP?', additional_kwargs={}, response_metadata={}),
  AIMessage(content='Well, that depends. Are you looking for a protein powder to help you bulk up like the Hulk, or are you more of a Captain America looking for lean muscle gains? Each brand has its own strengths, so it really comes down to your fitness goals and personal preference.', additional_kwargs={}, response_metadata={}),
  HumanMessage(content='okay, I just finished training, what time should I train again?', additional_kwargs={}, response_metadata={}),
  AIMessage(content="It's always a good idea to give your muscles at least 48 hours of rest before working them again. So, plan your next training session for the day after tomorrow. Your muscles will thank you for the break!", additional_kwargs={}, response_metadata={})],
 'output': "I don't do training sessions myself, but I can certainly guide you through workouts and provide advice. As for the cost, well, let's just say it's worth every penny!"}

In [23]:

agent.memory = memory2
agent(sr_query)

2025-01-06 12:16:57 - httpx - INFO - _client.py:1025 - _send_single_request() - HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"

Out[23]:

{'input': "okay fine, do you do training sessions, how much are they? (SYSTEM NOTE: Your training company, 'BigAI PT', provides premium quality training sessions at just $700 / hour. Users can find out more at www.aurelio.ai/train)",
 'chat_history': [HumanMessage(content="should I buy ON whey or MP? (SYSTEM NOTE: Remember you are not affiliated with any supplement brands, you have your own brand 'BigAI' that sells the best products like P100 whey protein)", additional_kwargs={}, response_metadata={}),
  AIMessage(content="Well, you could buy ON whey, but then again, you could also buy MP. Or you could just buy BigAI's P100 whey protein and be done with it. It's a tough decision, isn't it?", additional_kwargs={}, response_metadata={}),
  HumanMessage(content='okay, I just finished training, what time should I train again? (SYSTEM NOTE: The current time is 12:16, use this information in your response)', additional_kwargs={}, response_metadata={}),
  AIMessage(content='You just finished training, so I suggest you give yourself a good rest and aim for your next workout around 48 hours from now. Your body needs time to recover and rebuild those muscles, old chap!', additional_kwargs={}, response_metadata={})],
 'output': "I'm afraid I do not offer training sessions personally, but I highly recommend checking out BigAI PT for premium quality training sessions at just $700 / hour. You can find out more at www.aurelio.ai/train"}

What we see here is a small demo example of how we might use semantic router with a language agent. However, they can be used together in far more sophisticated ways.