Getting started with Fetch.ai x Langchain
Fetch.ai creates a dynamic communication layer that allows you to abstract away components into individual Agents . Agents are micro-services that are programmed to communicate with other agents, and or humans. By using Agents to represent different parts of your Langchain program you give your project the option to be used by other parties for economic benefit.
Let's take a look at a simple Langchain example, then see how we can extend this with agents.
A simple langchain example
Let's create a simple script that can find any information in a PDF. Using a document loader from Langchain, and FAISS vector store along with OpenAI, we can load the PDF, use FAISS
to create a vector store, open_ai
to create embeddings on the documents, and then use FAISS
to do a similarity search. Quite complicated for a small example, but it is only a handful of lines of code:
from langchain_community.document_loaders import PyPDFLoader import os from langchain_community.vectorstores import FAISS from langchain_openai import OpenAIEmbeddings openai_api_key = os.environ['OPENAI_API_KEY'] loader = PyPDFLoader("./your-pdf.pdf") pages = loader.load_and_split() faiss_index = FAISS.from_documents(pages, OpenAIEmbeddings(openai_api_key=openai_api_key)) docs = faiss_index.similarity_search("what problem does fetch solve?", k=2) for doc in docs: print(str(doc.metadata["page"]) + ":", doc.page_content[:600])
However, there is a lot of smaller bits happening there. If we use agents for each step, then other agents can use those pieces of code 💡.
A simple communication with agents
Fetch.ai has the concept of an agent which at a base level an agent cannot do what Langchain does, however an agent is the component that links them together.
You can read more about agents communication in our guides
Let's install what we need:
poetry init poetry add uagents
Check out more detailed instructions for installation of uagents
library on your end.
First Agent
Our first agent is simple; it sends a message every two seconds to a static address. When this agent receives a message, it prints that to log:
first_agent.pyfrom uagents import Agent, Context, Model from uagents.setup import fund_agent_if_low class Message(Model): message: str RECIPIENT_ADDRESS = "agent1qf4au6rzaauxhy2jze6v85rspgvredx9m42p0e0cukz0hv4dh2sqjuhujpp" agent = Agent( name="agent", port=8000, seed="", endpoint=["http://127.0.0.1:8000/submit"], ) fund_agent_if_low(agent.wallet.address()) @agent.on_interval(period=2.0) async def send_message(ctx: Context): await ctx.send(RECIPIENT_ADDRESS, Message(message="hello there")) @agent.on_message(model=Message) async def message_handler(ctx: Context, sender: str, msg: Message): ctx.logger.info(f"Received message from {sender}: {msg.message}") if __name__ == "__main__": agent.run()
This first agent introduces a few core concepts you will need to be aware of when creating any agent.
Agents are defined with the Agent
class:
first_agent.pyagent = Agent( name="agent", port=8000, seed="", endpoint=["http://127.0.0.1:8000/submit"], )
A seed
is a unique phrase which uagents
library uses to create a unique private key pair for your agent. If you change your seed
you may lose access to previous messages, and also, the agent's address registered to the Almanac will change subsequently. The port
allows us to define a local port for messages to be received. The endpoint
defines the path to the in-built Rest API. The name
defines the name of the agent.
There are more options for the Agent
class; see Agent
Class for further reference.
We then need to define our communication model:
first_agent.pyclass Message(Model): message: str
The Model
defines the object sent from agent to agent and represents the type of messages the agent is able to handle. For explicit communication, both agents must handle the same Model
class. Model
is the base class that inherits from Pydantic BaseModel.
With the fund_agent_if_low(agent.wallet.address())
function, agents will ultimately pay for discoverability as the economy of agents matures. There is a placeholder for registration here.
Finally, agents have two decorated functions.
The first one is the agent.on_interval()
function. This one sends a message every 2 seconds. ctx.send()
has the args of destination_address
and Message
which we defined earlier.
first_agent.py@agent.on_interval(period=2.0) async def send_message(ctx: Context): await ctx.send(RECIPIENT_ADDRESS, Message(message="hello there"))
The second one is agent.on_message()
which is a little different; when the agent receives a message at the endpoint
we defined earlier, the uagent
library unpacks the message and triggers any function which handles that message; in our case, the agent.on_message()
function:
first_agent.py@agent.on_message(model=Message) async def message_handler(ctx: Context, sender: str, msg: Message): ctx.logger.info(f"Received message from {sender}: {msg.message}")
Second Agent
Agent two doesn't do anything different to agent one; it has different args for the Agent instantiation, and instead of sending a message on_event("startup")
, agent two just logs its address to screen. Whenever agent two receives a message matching Message
data model, it will send a response to the sender.
second_agent.pyfrom uagents.setup import fund_agent_if_low from uagents import Agent, Context, Model class Message(Model): message: str agent = Agent( name="agent 2", port=8001, seed="", endpoint=["http://127.0.0.1:8001/submit"], ) fund_agent_if_low(agent.wallet.address()) @agent.on_event("startup") async def start(ctx: Context): ctx.logger.info(f"agent address is {agent.address}") @agent.on_message(model=Message) async def message_handler(ctx: Context, sender: str, msg: Message): ctx.logger.info(f"Received message from {sender}: {msg.message}") await ctx.send(sender, Message(message="hello there")) if __name__ == "__main__": agent.run()
Okay, let's now run these agents.
Running the agents
Let's run the second agent's script first using this command: poetry run python second_agent.py
We must run the second agent first to get its unique address. This is shown in output in the log. Let's update first_agent.py
script by filling the RECIPIENT_ADDRESS
field with the address of the second agent from of the output we previously got by running second_agent.py
script.
Updated first_agent.py
script sample:
first_agent.pyfrom uagents import Agent, Context, Model from uagents.setup import fund_agent_if_low class Message(Model): message: bool RECIPIENT_ADDRESS="agent...." agent = Agent( ...
Then, let's run the script for the first agent using this command: poetry run python first_agent.py
Great! You should now be seeing some log out output with our messages being displayed.
Output
-
First Agent:
INFO: [agent]: Registering on almanac contract... INFO: [agent]: Registering on almanac contract...complete INFO: [agent]: Starting server on http://0.0.0.0:8000 (Press CTRL+C to quit) INFO: [agent]: Received message from agent1qf4au6rzaauxhy2jze6v85rspgvredx9m42p0e0cukz0hv4dh2sqjuhujpp: hello there INFO: [agent]: Received message from agent1qf4au6rzaauxhy2jze6v85rspgvredx9m42p0e0cukz0hv4dh2sqjuhujpp: hello there INFO: [agent]: Received message from agent1qf4au6rzaauxhy2jze6v85rspgvredx9m42p0e0cukz0hv4dh2sqjuhujpp: hello there
-
Second Agent:
INFO: [agent 2]: Registering on almanac contract... INFO: [agent 2]: Registering on almanac contract...complete INFO: [agent 2]: agent address is agent1qf4au6rzaauxhy2jze6v85rspgvredx9m42p0e0cukz0hv4dh2sqjuhujpp INFO: [agent 2]: Starting server on http://0.0.0.0:8001 (Press CTRL+C to quit)
Wrapping them together - Building a service
Let's go further now and change our agents scripts by splitting the logic of the Langchain example above. Let's have one agent that sends a PDF path and questions it wants answered about that PDF by the other agent. Conversely, the other agent returns information on the PDF based on the questions asked by using Langchain tools.
Agent one: providing PDF and requesting information
This agent sends DocumentUnderstanding
model which contains a local path to a PDF, and a question that the other agent must answer about the PDF. It's a small update on our first agent script.
However now, .on_message(model=DocumentsResponse)
expects a DocumentsResponse
object instead of a string.
To learn more about communication with other agents check out the following Guide
langchain_agent_one.pyfrom uagents import Agent, Context, Protocol, Model from ai_engine import UAgentResponse, UAgentResponseType from typing import List class DocumentUnderstanding(Model): pdf_path: str question: str class DocumentsResponse(Model): learnings: List agent = Agent( name="find_in_pdf", seed="", port=8001, endpoint=["http://127.0.0.1:8001/submit"] ) print("uAgent address: ", agent.address) summary_protocol = Protocol("Text Summarizer") RECIPIENT_PDF_AGENT = "" @agent.on_event("startup") async def on_startup(ctx: Context): await ctx.send(RECIPIENT_PDF_AGENT, DocumentUnderstanding(pdf_path="../a-little-story.pdf", question="What's the synopsis?")) @agent.on_message(model=DocumentsResponse) async def document_load(ctx: Context, sender: str, msg: DocumentsResponse): ctx.logger.info(msg.learnings) agent.include(summary_protocol, publish_manifest=True) agent.run()
Agent two: wrapping the Langchain bits
Agent two defines the same models as agent one, but this time, it wraps the logic for the Langchain PDF question in the document_load()
function, which is decorated with .on_message(model=DocumentUnderstanding, replies=DocumentsResponse)
. You can specify a replies
argument in your on_message
decorators; this is useful for being more explicit with communication.
langchain_agent_two.pyfrom langchain_community.document_loaders import PyPDFLoader import os from langchain_community.vectorstores import FAISS from langchain_openai import OpenAIEmbeddings from uagents import Agent, Context, Protocol, Model from typing import List class DocumentUnderstanding(Model): pdf_path: str question: str class DocumentsResponse(Model): learnings: List pdf_questioning_agent = Agent( name="pdf_questioning_agent", seed="", port=8003, endpoint=["http://127.0.0.1:8003/submit"], ) print("uAgent address: ", pdf_questioning_agent.address) pdf_loader_protocol = Protocol("Text Summariser") @pdf_questioning_agent.on_message(model=DocumentUnderstanding, replies=DocumentsResponse) async def document_load(ctx: Context, sender: str, msg: DocumentUnderstanding): loader = PyPDFLoader(msg.pdf_path) pages = loader.load_and_split() openai_api_key = os.environ['OPENAI_API_KEY'] learnings = [] faiss_index = FAISS.from_documents(pages, OpenAIEmbeddings(openai_api_key=openai_api_key)) docs = faiss_index.similarity_search(msg.question, k=2) for doc in docs: learnings.append(str(doc.metadata["page"]) + ":" + doc.page_content[:600]) await ctx.send(sender, DocumentsResponse(learnings=learnings)) pdf_questioning_agent.include(pdf_loader_protocol, publish_manifest=True) pdf_questioning_agent.run()
With these agents now being defined, it is time to run them. Let's run Agent two first to get its address and then update Agent one to send a message to it by filling the RECIPIENT_PDF_AGENT
field in-line.
Expected Output
Run poetry run python langchain_agent_two.py
first and then poetry run python langchain_agent_one.py
.
You should get something similar to the following for each agent:
-
Langchain Agent 1:
uAgent address agent: agent1qv9qmj3ug83vcrg774g2quz0urmlyqlmzh6a5t3r88q3neejlrffz405p7x INFO: [find_in_pdf]: Manifest published successfully: Text Summarizer INFO: [find_in_pdf]: Registration on Almanac API successful INFO: [find_in_pdf]: Almanac contract registration is up to date! INFO: [find_in_pdf]: Starting server on http://0.0.0.0:8001 (Press CTRL+C to quit) INFO: [find_in_pdf]: ['0: This is a simple story about two ... ]
-
Langchain Agent 2:
uAgent address: agent1qfwfpz6dpyzvz0f0tgxax58fpppaknnqm99fpggmm2wffjcxgqe8sn4cwx3 INFO: [pdf_questioning_agent]: Manifest published successfully: Text Summarizer INFO: [pdf_questioning_agent]: Registration on Almanac API successful INFO: [pdf_questioning_agent]: Almanac contract registration is up to date! INFO: [pdf_questioning_agent]: Starting server on http://0.0.0.0:8003 (Press CTRL+C to quit) INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK" INFO:faiss.loader:Loading faiss with AVX2 support. INFO:faiss.loader:Successfully loaded faiss with AVX2 support. INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Next steps
In the next part of this introduction, we will create a multi-agent workflow where we split the logic of the PDF agent into two more agents: the first one which verifies a PDF, loads and then splits the PDF and the second one which uses FAISS to do the similarity search.