Initially, when ChatGPT simply appeared, we used easy prompts to get solutions to our questions. Then, we encountered points with hallucinations and started utilizing RAG (Retrieval Augmented Era) to supply extra context to LLMs. After that, we began experimenting with AI brokers, the place LLMs act as a reasoning engine and may resolve what to do subsequent, which instruments to make use of, and when to return the ultimate reply.
The following evolutionary step is to create groups of such brokers that may collaborate with one another. This method is logical because it mirrors human interactions. We work in groups the place every member has a particular position:
- The product supervisor proposes the subsequent undertaking to work on.
- The designer creates its appear and feel.
- The software program engineer develops the answer.
- The analyst examines the information to make sure it performs as anticipated and identifies methods to enhance the product for purchasers.
Equally, we are able to create a crew of AI brokers, every specializing in one area. They’ll collaborate and attain a last conclusion collectively. Simply as specialization enhances efficiency in actual life, it may additionally profit the efficiency of AI brokers.
One other benefit of this method is elevated flexibility. Every agent can function with its personal immediate, set of instruments and even LLM. For example, we are able to use completely different fashions for various elements of our system. You should utilize GPT-4 for the agent that wants extra reasoning and GPT-3.5 for the one which does solely easy extraction. We are able to even fine-tune the mannequin for small particular duties and use it in our crew of brokers.
The potential drawbacks of this method are time and price. A number of interactions and information sharing between brokers require extra calls to LLM and eat further tokens. This might end in longer wait instances and elevated bills.
There are a number of frameworks out there for multi-agent methods at this time.
Listed below are among the hottest ones:
- AutoGen: Developed by Microsoft, AutoGen makes use of a conversational method and was one of many earliest frameworks for multi-agent methods,
- LangGraph: Whereas not strictly a multi-agent framework, LangGraph permits for outlining advanced interactions between actors utilizing a graph construction. So, it may also be tailored to create multi-agent methods.
- CrewAI: Positioned as a high-level framework, CrewAI facilitates the creation of “crews” consisting of role-playing brokers able to collaborating in varied methods.
I’ve determined to begin experimenting with multi-agent frameworks from CrewAI because it’s fairly extensively well-liked and person pleasant. So, it appears like possibility to start with.
On this article, I’ll stroll you thru how you can use CrewAI. As analysts, we’re the area specialists answerable for documenting varied information sources and addressing associated questions. We’ll discover how you can automate these duties utilizing multi-agent frameworks.
Let’s begin with organising the setting. First, we have to set up the CrewAI primary bundle and an extension to work with instruments.
pip set up crewai
pip set up 'crewai[tools]'
CrewAI was developed to work primarily with OpenAI API, however I’d additionally prefer to attempt it with an area mannequin. In line with the ChatBot Area Leaderboard, the very best mannequin you possibly can run in your laptop computer is Llama 3 (8b parameters). It is going to be probably the most possible possibility for our use case.
We are able to entry Llama fashions utilizing Ollama. Set up is fairly simple. It’s essential to obtain Ollama from the web site after which undergo the set up course of. That’s it.
Now, you possibly can take a look at the mannequin in CLI by working the next command.
ollama run llama3
For instance, you possibly can ask one thing like this.
Let’s create a customized Ollama mannequin to make use of later in CrewAI.
We’ll begin with a ModelFile (documentation). I solely specified the bottom mannequin (llama3), temperature and cease sequence. Nonetheless, you may add extra options. For instance, you possibly can decide the system message utilizing SYSTEM key phrase.
FROM llama3# set parameters
PARAMETER temperature 0.5
PARAMETER cease Consequence
I’ve saved it right into a Llama3ModelFile file.
Let’s create a bash script to load the bottom mannequin for Ollama and create the customized mannequin we outlined in ModelFile.
#!/bin/zsh# outline variables
model_name="llama3"
custom_model_name="crewai-llama3"
# load the bottom mannequin
ollama pull $model_name
# create the mannequin file
ollama create $custom_model_name -f ./Llama3ModelFile
Let’s execute this file.
chmod +x ./llama3_setup.sh
./llama3_setup.sh
You could find each recordsdata on GitHub: Llama3ModelFile and llama3_setup.sh
We have to initialise the next environmental variables to make use of the native Llama mannequin with CrewAI.
os.environ["OPENAI_API_BASE"]='http://localhost:11434/v1'os.environ["OPENAI_MODEL_NAME"]='crewai-llama3'
# custom_model_name from the bash script
os.environ["OPENAI_API_KEY"] = "NA"
We’ve completed the setup and are able to proceed our journey.
As analysts, we frequently play the position of material specialists for information and a few data-related instruments. In my earlier crew, we used to have a channel with nearly 1K contributors, the place we have been answering numerous questions on our information and the ClickHouse database we used as storage. It took us numerous time to handle this channel. It could be fascinating to see whether or not such duties may be automated with LLMs.
For this instance, I’ll use the ClickHouse database. In case you’re , You’ll be able to study extra about ClickHouse and how you can set it up regionally in my earlier article. Nonetheless, we received’t utilise any ClickHouse-specific options, so be at liberty to stay to the database you recognize.
I’ve created a fairly easy information mannequin to work with. There are simply two tables in our DWH (Information Warehouse): ecommerce_db.customers and ecommerce_db.classes. As you may guess, the primary desk accommodates details about the customers of our service.
The ecommerce_db.classes desk shops details about person classes.
Concerning information supply administration, analysts sometimes deal with duties like writing and updating documentation and answering questions on this information. So, we’ll use LLM to jot down documentation for the desk within the database and train it to reply questions on information or ClickHouse.
However earlier than transferring on to the implementation, let’s study extra in regards to the CrewAI framework and its core ideas.
The cornerstone of a multi-agent framework is an agent idea. In CrewAI, brokers are powered by role-playing. Function-playing is a tactic whenever you ask an agent to undertake a persona and behave like a top-notch backend engineer or useful buyer assist agent. So, when making a CrewAI agent, it’s essential to specify every agent’s position, aim, and backstory in order that LLM is aware of sufficient to play this position.
The brokers’ capabilities are restricted with out instruments (features that brokers can execute and get outcomes). With CrewAI, you should use one of many predefined instruments (for instance, to go looking the Web, parse a web site, or do RAG on a doc), create a customized device your self or use LangChain instruments. So, it’s fairly simple to create a strong agent.
Let’s transfer on from brokers to the work they’re doing. Brokers are engaged on duties (particular assignments). For every process, we have to outline an outline, anticipated output (definition of executed), set of accessible instruments and assigned agent. I actually like that these frameworks comply with the managerial finest practices like a transparent definition of executed for the duties.
The following query is how you can outline the execution order for duties: which one to work on first, which of them can run in parallel, and many others. CrewAI carried out processes to orchestrate the duties. It gives a few choices:
- Sequential —probably the most simple method when duties are known as one after one other.
- Hierarchical — when there’s a supervisor (specified as LLM mannequin) that creates and delegates duties to the brokers.
Additionally, CrewAI is engaged on a consensual course of. In such a course of, brokers will have the ability to make selections collaboratively with a democratic method.
There are different levers you should use to tweak the method of duties’ execution:
- You’ll be able to mark duties as “asynchronous”, then they are going to be executed in parallel, so it is possible for you to to get a solution sooner.
- You should utilize the “human enter” flag on a process, after which the agent will ask for human approval earlier than finalising the output of this process. It will possibly permit you to add an oversight to the method.
We’ve outlined all the first constructing blocks and may focus on the holly grail of CrewAI — crew idea. The crew represents the crew of brokers and the set of duties they are going to be engaged on. The method for collaboration (processes we mentioned above) may also be outlined on the crew stage.
Additionally, we are able to arrange the reminiscence for a crew. Reminiscence is essential for environment friendly collaboration between the brokers. CrewAI helps three ranges of reminiscence:
- Brief-term reminiscence shops info associated to the present execution. It helps brokers to work collectively on the present process.
- Lengthy-term reminiscence is information in regards to the earlier executions saved within the native database. One of these reminiscence permits brokers to study from earlier iterations and enhance over time.
- Entity reminiscence captures and buildings details about entities (like personas, cities, and many others.)
Proper now, you possibly can solely swap on all kinds of reminiscence for a crew with none additional customisation. Nonetheless, it doesn’t work with the Llama fashions.
We’ve discovered sufficient in regards to the CrewAI framework, so it’s time to begin utilizing this information in apply.
Let’s begin with a easy process: placing collectively the documentation for our DWH. As we mentioned earlier than, there are two tables in our DWH, and I wish to create an in depth description for them utilizing LLMs.
First method
To start with, we want to consider the crew construction. Consider this as a typical managerial process. Who would you rent for such a job?
I’d break this process into two elements: retrieving information from a database and writing documentation. So, we want a database specialist and a technical author. The database specialist wants entry to a database, whereas the author received’t want any particular instruments.
Now, we have now a high-level plan. Let’s create the brokers.
For every agent, I’ve specified the position, aim and backstory. I’ve tried my finest to supply brokers with all of the wanted context.
database_specialist_agent = Agent(
position = "Database specialist",
aim = "Present information to reply enterprise questions utilizing SQL",
backstory = '''You might be an skilled in SQL, so you possibly can assist the crew
to collect wanted information to energy their selections.
You might be very correct and take note of all of the nuances in information.''',
allow_delegation = False,
verbose = True
)tech_writer_agent = Agent(
position = "Technical author",
aim = '''Write participating and factually correct technical documentation
for information sources or instruments''',
backstory = '''
You might be an skilled in each know-how and communications, so you possibly can simply clarify even refined ideas.
You base your work on the factual info offered by your colleagues.
Your texts are concise and may be simply understood by a large viewers.
You utilize skilled however somewhat a casual type in your communication.
''',
allow_delegation = False,
verbose = True
)
We’ll use a easy sequential course of, so there’s no want for brokers to delegate duties to one another. That’s why I specified allow_delegation = False.
The following step is setting the duties for brokers. However earlier than transferring to them, we have to create a customized device to connect with the database.
First, I put collectively a operate to execute ClickHouse queries utilizing HTTP API.
CH_HOST = 'http://localhost:8123' # default deal with def get_clickhouse_data(question, host = CH_HOST, connection_timeout = 1500):
r = requests.submit(host, params = {'question': question},
timeout = connection_timeout)
if r.status_code == 200:
return r.textual content
else:
return 'Database returned the next error:n' + r.textual content
When working with LLM brokers, it’s necessary to make instruments fault-tolerant. For instance, if the database returns an error (status_code != 200), my code received’t throw an exception. As a substitute, it’ll return the error description to the LLM so it could actually try to resolve the difficulty.
To create a CrewAI customized device, we have to derive our class from crewai_tools.BaseTool, implement the _run technique after which create an occasion of this class.
from crewai_tools import BaseToolclass DatabaseQuery(BaseTool):
title: str = "Database Question"
description: str = "Returns the results of SQL question execution"
def _run(self, sql_query: str) -> str:
# Implementation goes right here
return get_clickhouse_data(sql_query)
database_query_tool = DatabaseQuery()
Now, we are able to set the duties for the brokers. Once more, offering clear directions and all of the context to LLM is essential.
table_description_task = Process(
description = '''Present the excellent overview for the information
in desk {desk}, in order that it is simple to grasp the construction
of the information. This process is essential to place collectively the documentation
for our database''',
expected_output = '''The great overview of {desk} within the md format.
Embrace 2 sections: columns (record of columns with their varieties)
and examples (the primary 30 rows from desk).''',
instruments = [database_query_tool],
agent = database_specialist_agent
)table_documentation_task = Process(
description = '''Utilizing offered details about the desk,
put collectively the detailed documentation for this desk in order that
folks can use it in apply''',
expected_output = '''Nicely-written detailed documentation describing
the information scheme for the desk {desk} in markdown format,
that offers the desk overview in 1-2 sentences then then
describes every columm. Construction the columns description
as a markdown desk with column title, kind and outline.''',
instruments = [],
output_file="table_documentation.md",
agent = tech_writer_agent
)
You may need observed that I’ve used {desk} placeholder within the duties’ descriptions. We’ll use desk as an enter variable when executing the crew, and this worth will likely be inserted into all placeholders.
Additionally, I’ve specified the output file for the desk documentation process to avoid wasting the ultimate end result regionally.
We have now all we want. Now, it’s time to create a crew and execute the method, specifying the desk we’re desirous about. Let’s attempt it with the customers desk.
crew = Crew(
brokers = [database_specialist_agent, tech_writer_agent],
duties = [table_description_task, table_documentation_task],
verbose = 2
)end result = crew.kickoff({'desk': 'ecommerce_db.customers'})
It’s an thrilling second, and I’m actually wanting ahead to seeing the end result. Don’t fear if execution takes a while. Brokers make a number of LLM calls, so it’s completely regular for it to take a couple of minutes. It took 2.5 minutes on my laptop computer.
We requested LLM to return the documentation in markdown format. We are able to use the next code to see the formatted end in Jupyter Pocket book.
from IPython.show import Markdown
Markdown(end result)
At first look, it appears nice. We’ve bought the legitimate markdown file describing the customers’ desk.
However wait, it’s incorrect. Let’s see what information we have now in our desk.
The columns listed within the documentation are fully completely different from what we have now within the database. It’s a case of LLM hallucinations.
We’ve set verbose = 2 to get the detailed logs from CrewAI. Let’s learn by way of the execution logs to determine the foundation explanation for the issue.
First, the database specialist couldn’t question the database as a result of problems with quotes.
The specialist didn’t handle to resolve this drawback. Lastly, this chain has been terminated by CrewAI with the next output: Agent stopped as a result of iteration restrict or time restrict.
This implies the technical author didn’t obtain any factual details about the information. Nonetheless, the agent continued and produced fully faux outcomes. That’s how we ended up with incorrect documentation.
Fixing the problems
Although our first iteration wasn’t profitable, we’ve discovered lots. We have now (at the least) two areas for enchancment:
- Our database device is just too tough for the mannequin, and the agent struggles to make use of it. We are able to make the device extra tolerant by eradicating quotes from the start and finish of the queries. This answer is just not best since legitimate SQL can finish with a quote, however let’s attempt it.
- Our technical author isn’t basing its output on the enter from the database specialist. We have to tweak the immediate to focus on the significance of offering solely factual info.
So, let’s attempt to repair these issues. First, we’ll repair the device — we are able to leverage strip to get rid of quotes.
CH_HOST = 'http://localhost:8123' # default deal with def get_clickhouse_data(question, host = CH_HOST, connection_timeout = 1500):
r = requests.submit(host, params = {'question': question.strip('"').strip("'")},
timeout = connection_timeout)
if r.status_code == 200:
return r.textual content
else:
return 'Database returned the next error:n' + r.textual content
Then, it’s time to replace the immediate. I’ve included statements emphasizing the significance of sticking to the details in each the agent and process definitions.
tech_writer_agent = Agent(
position = "Technical author",
aim = '''Write participating and factually correct technical documentation
for information sources or instruments''',
backstory = '''
You might be an skilled in each know-how and communications, so that you
can simply clarify even refined ideas.
Your texts are concise and may be simply understood by huge viewers.
You utilize skilled however somewhat casual type in your communication.
You base your work on the factual info offered by your colleagues.
You follow the details within the documentation and use ONLY
info offered by the colleagues not including something.''',
allow_delegation = False,
verbose = True
)table_documentation_task = Process(
description = '''Utilizing offered details about the desk,
put collectively the detailed documentation for this desk in order that
folks can use it in apply''',
expected_output = '''Nicely-written detailed documentation describing
the information scheme for the desk {desk} in markdown format,
that offers the desk overview in 1-2 sentences then then
describes every columm. Construction the columns description
as a markdown desk with column title, kind and outline.
The documentation is predicated ONLY on the knowledge offered
by the database specialist with none additions.''',
instruments = [],
output_file = "table_documentation.md",
agent = tech_writer_agent
)
Let’s execute our crew as soon as once more and see the outcomes.
We’ve achieved a bit higher end result. Our database specialist was in a position to execute queries and examine the information, which is a major win for us. Moreover, we are able to see all of the related fields within the end result desk, although there are many different fields as effectively. So, it’s nonetheless not solely appropriate.
I as soon as once more seemed by way of the CrewAI execution log to determine what went incorrect. The problem lies in getting the record of columns. There’s no filter by database, so it returns some unrelated columns that seem within the end result.
SELECT column_name
FROM information_schema.columns
WHERE table_name = 'customers'
Additionally, after taking a look at a number of makes an attempt, I observed that the database specialist, every so often, executes choose * from <desk> question. It would trigger some points in manufacturing as it’d generate numerous information and ship it to LLM.
Extra specialised instruments
We are able to present our agent with extra specialised instruments to enhance our answer. Presently, the agent has a device to execute any SQL question, which is versatile and highly effective however liable to errors. We are able to create extra centered instruments, reminiscent of getting desk construction and top-N rows from the desk. Hopefully, it’ll cut back the variety of errors.
class TableStructure(BaseTool):
title: str = "Desk construction"
description: str = "Returns the record of columns and their varieties"def _run(self, desk: str) -> str:
desk = desk.strip('"').strip("'")
return get_clickhouse_data(
'describe {desk} format TabSeparatedWithNames'
.format(desk = desk)
)
class TableExamples(BaseTool):
title: str = "Desk examples"
description: str = "Returns the primary N rows from the desk"
def _run(self, desk: str, n: int = 30) -> str:
desk = desk.strip('"').strip("'")
return get_clickhouse_data(
'choose * from {desk} restrict {n} format TabSeparatedWithNames'
.format(desk = desk, n = n)
)
table_structure_tool = TableStructure()
table_examples_tool = TableExamples()
Now, we have to specify these instruments within the process and re-run our script. After the primary try, I bought the next output from the Technical Author.
Process output: This last reply gives an in depth and factual description
of the ecommerce_db.customers desk construction, together with column names, varieties,
and descriptions. The documentation adheres to the offered info
from the database specialist with none additions or modifications.
Extra centered instruments helped the database specialist retrieve the right desk info. Nonetheless, regardless that the author had all the required info, we didn’t get the anticipated end result.
As we all know, LLMs are probabilistic, so I gave it one other attempt. And hooray, this time, the end result was fairly good.
It’s not excellent because it nonetheless consists of some irrelevant feedback and lacks the general description of the desk. Nonetheless, offering extra specialised instruments has undoubtedly paid off. It additionally helped to stop points when the agent tried to load all the information from the desk.
High quality assurance specialist
We’ve achieved fairly good outcomes, however let’s see if we are able to enhance them additional. A typical apply in multi-agent setups is high quality assurance, which provides the ultimate evaluation stage earlier than finalising the outcomes.
Let’s create a brand new agent — a High quality Assurance Specialist, who will likely be answerable for evaluation.
qa_specialist_agent = Agent(
position = "High quality Assurance specialist",
aim = """Guarantee the best high quality of the documentation we offer
(that it is appropriate and straightforward to grasp)""",
backstory = '''
You're employed as a High quality Assurance specialist, checking the work
from the technical author and guaranteeing that it is inline
with our highest requirements.
It's essential to verify that the technical author gives the total full
solutions and make no assumptions.
Additionally, it's essential to be sure that the documentation addresses
all of the questions and is simple to grasp.
''',
allow_delegation = False,
verbose = True
)
Now, it’s time to explain the evaluation process. I’ve used the context parameter to specify that this process requires outputs from each table_description_task and table_documentation_task.
qa_review_task = Process(
description = '''
Evaluate the draft documentation offered by the technical author.
Be certain that the documentation absolutely solutions all of the questions:
the aim of the desk and its construction within the type of desk.
Be sure that the documentation is per the knowledge
offered by the database specialist.
Double verify that there are not any irrelevant feedback within the last model
of documentation.
''',
expected_output = '''
The ultimate model of the documentation in markdown format
that may be printed.
The documentation ought to absolutely deal with all of the questions, be constant
and comply with our skilled however casual tone of voice.
''',
instruments = [],
context = [table_description_task, table_documentation_task],
output_file="checked_table_documentation.md",
agent = qa_specialist_agent
)
Let’s replace our crew and run it.
full_crew = Crew(
brokers=[database_specialist_agent, tech_writer_agent, qa_specialist_agent],
duties=[table_description_task, table_documentation_task, qa_review_task],
verbose = 2,
reminiscence = False # do not work with Llama
)full_result = full_crew.kickoff({'desk': 'ecommerce_db.customers'})
We now have extra structured and detailed documentation due to the addition of the QA stage.
Delegation
With the addition of the QA specialist, it might be fascinating to check the delegation mechanism. The QA specialist agent may need questions or requests that it may delegate to different brokers.
I attempted utilizing the delegation with Llama 3, nevertheless it didn’t go effectively. Llama 3 struggled to name the co-worker device accurately. It couldn’t specify the right co-worker’s title.
We achieved fairly good outcomes with an area mannequin that may run on any laptop computer, however now it’s time to change gears and use a far more highly effective mannequin — GPT-4o.
To do it, we simply have to replace the next setting variables.
os.environ["OPENAI_MODEL_NAME"] = 'gpt-4o'
os.environ["OPENAI_API_KEY"] = config['OPENAI_API_KEY'] # your OpenAI key
To change on the delegation, we should always specify allow_delegation = True for the QA specialist agent.
Additionally, we are able to use helpful reminiscence performance for OpenAI fashions, as talked about above. The reminiscence will enable our brokers to share their information with one another throughout execution and leverage long-term reminiscence to get info from earlier executions.
full_crew = Crew(
brokers = [database_specialist_agent, tech_writer_agent, qa_specialist_agent],
duties = [table_description_task, table_documentation_task, qa_review_task],
verbose = 2,
reminiscence = True
)full_result = full_crew.kickoff({'desk': 'ecommerce_db.customers'})
Let’s see the CrewAI execution logs to grasp how delegation works. Listed below are all of the logs for the QA specialist. We are able to see that it reached out to the database specialist to double-check the knowledge.
Thought: I have to evaluation the offered draft documentation to make sure it
absolutely solutions all of the questions and is full, appropriate, and straightforward
to grasp. Particularly, I would like to make sure that it explains the aim
of the desk, its construction, and consists of related examples with out
any irrelevant feedback.First, I'll evaluate the 2 drafts to determine any discrepancies and
lacking info. Then, I'll confirm the accuracy and completeness of
the knowledge offered.
**Comparability:**
1. **Introduction:**
- First draft: Contains an introduction and examples.
- Second draft: Provides a extra detailed introduction in regards to the goal and significance of the desk.
2. **Columns:**
- Each drafts have the identical column construction and descriptions.
3. **Extra Data:**
- Each drafts embrace related further info, however the second draft is extra concise.
4. **Examples:**
- The primary draft consists of examples, that are lacking within the second draft.
5. **Ultimate Checks:**
- Guarantee there are not any irrelevant feedback.
- Confirm consistency with info offered by the database specialist.
Motion: Ask query to co-worker
Motion Enter: 88
<-- omitted logs from the database specialist -->
Based mostly on the offered context and examples for the `ecommerce_db.customers` desk,
the column descriptions and examples seem correct. Listed below are the main points:
- `user_id`: This column accommodates a singular identifier for every person.
The offered examples (e.g., 1000001, 1000002) align with this description
as they're distinctive numeric identifiers.
- `nation`: This column specifies the nation of the person. The offered
examples (e.g., United Kingdom, France, Germany, Netherlands) are
all legitimate nation names.
- `is_active`: This column signifies whether or not the person is energetic (1) or not (0). The offered examples (e.g., 0, 1) accurately characterize this binary standing.
- `age`: This column exhibits the age of the person. The offered examples (e.g., 70,
87, 88, 25, 48, 78, 65, 31, 66, 73) are all legitimate ages.
Due to this fact, the column descriptions and examples offered for the
`ecommerce_db.customers` desk are correct and meet the anticipated standards.
Thought: Based mostly on the commentary from the database specialist,
the column descriptions and examples offered for the `ecommerce_db.customers`
desk are correct. Now, I'll consolidate the very best parts from each drafts,
guaranteeing the documentation is full, clear, and freed from irrelevant feedback.
Motion: Finalize the documentation in markdown format, incorporating
the detailed introduction, examples, and guaranteeing the tone is skilled
however casual.
Once I tried the delegation for the primary time, I didn’t allow reminiscence, which led to incorrect outcomes. The information specialist and the technical author initially returned the right info. Nonetheless, when the QA specialist returned with the follow-up questions, they began to hallucinate. So, it appears like delegation works higher when reminiscence is enabled.
Right here’s the ultimate output from GPT-4o. The end result appears fairly good now. We undoubtedly can use LLMs to automate documentation.
So, the primary process has been solved!
I used the identical script to generate documentation for the ecommerce_db.classes desk as effectively. It is going to be helpful for our subsequent process. So, let’s not waste any time and transfer on.
Our subsequent process is answering questions primarily based on the documentation because it’s frequent for a lot of information analysts (and different specialists).
We’ll begin easy and can create simply two brokers:
- The documentation assist specialist will likely be answering questions primarily based on the docs,
- The assist QA agent will evaluation the reply earlier than sharing it with the shopper.
We might want to empower the documentation specialist with a few instruments that can enable them to see all of the recordsdata saved within the listing and browse the recordsdata. It’s fairly simple since CrewAI has carried out such instruments.
from crewai_tools import DirectoryReadTool, FileReadTooldocumentation_directory_tool = DirectoryReadTool(
listing = '~/crewai_project/ecommerce_documentation')
base_file_read_tool = FileReadTool()
Nonetheless, since Llama 3 retains fighting quotes when calling instruments, I needed to create a customized device on high of the FileReaderTool to beat this problem.
from crewai_tools import BaseToolclass FileReadToolUPD(BaseTool):
title: str = "Learn a file's content material"
description: str = "A device that can be utilized to learn a file's content material."
def _run(self, file_path: str) -> str:
# Implementation goes right here
return base_file_read_tool._run(file_path = file_path.strip('"').strip("'"))
file_read_tool = FileReadToolUPD()
Subsequent, as we did earlier than, we have to create brokers, duties and crew.
data_support_agent = Agent(
position = "Senior Information Help Agent",
aim = "Be probably the most useful assist for you colleagues",
backstory = '''You're employed as a assist for data-related questions
within the firm.
Although you are an enormous skilled in our information warehouse, you double verify
all of the details in documentation.
Our documentation is completely up-to-date, so you possibly can absolutely depend on it
when answering questions (you need not verify the precise information
in database).
Your work is essential for the crew success. Nonetheless, keep in mind
that examples of desk rows do not present all of the attainable values.
It's essential to be sure that you present the very best assist: answering
all of the questions, making no assumptions and sharing solely the factual information.
Be artistic attempt your finest to resolve the shopper drawback.
''',
allow_delegation = False,
verbose = True
)qa_support_agent = Agent(
position = "Help High quality Assurance Agent",
aim = """Guarantee the best high quality of the solutions we offer
to the purchasers""",
backstory = '''You're employed as a High quality Assurance specialist, checking the work
from assist brokers and guaranteeing that it is inline with our highest requirements.
It's essential to verify that the agent gives the total full solutions
and make no assumptions.
Additionally, it's essential to be sure that the documentation addresses all
the questions and is simple to grasp.
''',
allow_delegation = False,
verbose = True
)
draft_data_answer = Process(
description = '''Essential buyer {buyer} reached out to you
with the next query:
```
{query}
```
Your process is to supply the very best reply to all of the factors within the query
utilizing all out there info and never making any assumprions.
If you do not have sufficient info to reply the query, simply say
that you do not know.''',
expected_output = '''The detailed informative reply to the shopper's
query that addresses all the purpose talked about.
Be sure that reply is full and stict to details
(with none further info not primarily based on the factual information)''',
instruments = [documentation_directory_tool, file_read_tool],
agent = data_support_agent
)
answer_review = Process(
description = '''
Evaluate the draft reply offered by the assist agent.
Be certain that the it absolutely solutions all of the questions talked about
within the preliminary inquiry.
Be sure that the reply is constant and would not embrace any assumptions.
''',
expected_output = '''
The ultimate model of the reply in markdown format that may be shared
with the shopper.
The reply ought to absolutely deal with all of the questions, be constant
and comply with our skilled however casual tone of voice.
We're very chill and pleasant firm, so do not forget to incorporate
all of the well mannered phrases.
''',
instruments = [],
agent = qa_support_agent
)
qna_crew = Crew(
brokers = [data_support_agent, qa_support_agent],
duties = [draft_data_answer, answer_review],
verbose = 2,
reminiscence = False # do not work with Llama
)
Let’s see the way it works in apply.
end result = qna_crew.kickoff(
{'buyer': "Max",
'query': """Hey crew, I hope you are doing effectively. I would like to search out
the numbers earlier than our CEO presentation tomorrow, so I'll actually
admire your assist.
I have to calculate the variety of classes from our Home windows customers in 2023. I've tried to search out the desk with such information in our information warehouse, however wasn't in a position to.
Do you've gotten any concepts whether or not we retailer the wanted information someplace,
in order that I can question it? """
}
)
We’ve bought a well mannered, sensible and useful reply in return. That’s actually nice.
**Whats up Max,**Thanks for reaching out together with your query! I am pleased that can assist you
discover the variety of classes from Home windows customers in 2023.
After reviewing our documentation, I discovered that we do retailer information
associated to classes and customers in our ecommerce database, particularly in
the `ecommerce_db.classes` desk.
To reply your query, I can offer you a step-by-step information
on how you can question this desk utilizing SQL. First, you should use the `session_id`
column together with the `os` column filtering for "Home windows" and
the `action_date` column filtering for dates in 2023.
Then, you possibly can group the outcomes by `os` utilizing the `GROUP BY` clause
to rely the variety of classes that meet these situations.
Here is a pattern SQL question that ought to provide the desired output:
```sql
SELECT COUNT(*)
FROM ecommerce_db.classes
WHERE os = 'Home windows'
AND action_date BETWEEN '2023-01-01' AND '2023-12-31'
GROUP BY os;
```
This question will return the overall variety of classes from Home windows
customers in 2023. I hope this helps! If in case you have any additional questions or
want extra help, please do not hesitate to ask.
Let’s complicate the duty a bit. Suppose we are able to get not solely questions on our information but additionally about our device (ClickHouse). So, we could have one other agent within the crew — ClickHouse Guru. To provide our CH agent some information, I’ll share a documentation web site with it.
from crewai_tools import ScrapeWebsiteTool, WebsiteSearchTool
ch_documenation_tool = ScrapeWebsiteTool(
'https://clickhouse.com/docs/en/guides/creating-tables')
If it’s essential to work with a prolonged doc, you may attempt utilizing RAG (Retrieval Augmented era) — WebsiteSearchTool. It’s going to calculate embeddings and retailer them regionally in ChromaDB. In our case, we’ll follow a easy web site scraper device.
Now that we have now two material specialists, we have to resolve who will likely be engaged on the questions. So, it’s time to make use of a hierarchical course of and add a supervisor to orchestrate all of the duties.
CrewAI gives the supervisor implementation, so we solely have to specify the LLM mannequin. I’ve picked the GPT-4o.
from langchain_openai import ChatOpenAI
from crewai import Course ofcomplext_qna_crew = Crew(
brokers = [ch_support_agent, data_support_agent, qa_support_agent],
duties = [draft_ch_answer, draft_data_answer, answer_review],
verbose = 2,
manager_llm = ChatOpenAI(mannequin='gpt-4o', temperature=0),
course of = Course of.hierarchical,
reminiscence = False
)
At this level, I needed to swap from Llama 3 to OpenAI fashions once more to run a hierarchical course of because it hasn’t labored for me with Llama (just like this problem).
Now, we are able to attempt our new crew with several types of questions (both associated to our information or ClickHouse database).
ch_result = complext_qna_crew.kickoff(
{'buyer': "Maria",
'query': """Good morning, crew. I am utilizing ClickHouse to calculate
the variety of prospects.
Might you please remind whether or not there's an possibility so as to add totals
in ClickHouse?"""
}
)doc_result = complext_qna_crew.kickoff(
{'buyer': "Max",
'query': """Hey crew, I hope you are doing effectively. I would like to search out
the numbers earlier than our CEO presentation tomorrow, so I'll actually
admire your assist.
I have to calculate the variety of classes from our Home windows customers
in 2023. I've tried to search out the desk with such information
in our information warehouse, however wasn't in a position to.
Do you've gotten any concepts whether or not we retailer the wanted information someplace,
in order that I can question it. """
}
)
If we take a look at the ultimate solutions and logs (I’ve omitted them right here since they’re fairly prolonged, but yow will discover them and full logs on GitHub), we’ll see that the supervisor was in a position to orchestrate accurately and delegate duties to co-workers with related information to deal with the shopper’s query. For the primary (ClickHouse-related) query, we bought an in depth reply with examples and attainable implications of utilizing WITH TOTALS performance. For the data-related query, fashions returned roughly the identical info as we’ve seen above.
So, we’ve constructed a crew that may reply varied kinds of questions primarily based on the documentation, whether or not from an area file or a web site. I believe it’s a superb end result.
You could find all of the code on GitHub.
On this article, we’ve explored utilizing the CrewAI multi-agent framework to create an answer for writing documentation primarily based on tables and answering associated questions.
Given the intensive performance we’ve utilised, it’s time to summarise the strengths and weaknesses of this framework.
General, I discover CrewAI to be an extremely helpful framework for multi-agent methods:
- It’s simple, and you’ll construct your first prototype shortly.
- Its flexibility permits to resolve fairly refined enterprise issues.
- It encourages good practices like role-playing.
- It gives many helpful instruments out of the field, reminiscent of RAG and a web site parser.
- The assist of several types of reminiscence enhances the brokers’ collaboration.
- Constructed-in guardrails assist stop brokers from getting caught in repetitive loops.
Nonetheless, there are areas that might be improved:
- Whereas the framework is easy and straightforward to make use of, it’s not very customisable. For example, you at present can’t create your personal LLM supervisor to orchestrate the processes.
- Typically, it’s fairly difficult to get the total detailed info from the documentation. For instance, it’s clear that CrewAI carried out some guardrails to stop repetitive operate calls, however the documentation doesn’t absolutely clarify the way it works.
- One other enchancment space is transparency. I like to grasp how frameworks work underneath the hood. For instance, in Langchain, you should use
langchain.debug = Trueto see all of the LLM calls. Nonetheless, I haven’t found out how you can get the identical stage of element with CrewAI. - The total assist for the native fashions can be a terrific addition, as the present implementation both lacks some options or is tough to get working correctly.
The area and instruments for LLMs are evolving quickly, so I’m hopeful that we’ll see a number of progress within the close to future.
Thank you numerous for studying this text. I hope this text was insightful for you. If in case you have any follow-up questions or feedback, please go away them within the feedback part.
This text is impressed by the “Multi AI Agent Methods with CrewAI” quick course from DeepLearning.AI.
