Agentic AI 102: Guardrails and Agent Evaluation

Within the first submit of this collection (Agentic AI 101: Starting Your Journey Building AI Agents), we talked in regards to the fundamentals of making AI Brokers and launched ideas like reasoning, reminiscence, and instruments.

After all, that first submit touched solely the floor of this new space of the info trade. There’s a lot extra that may be achieved, and we’re going to study extra alongside the best way on this collection.

So, it’s time to take one step additional.

On this submit, we are going to cowl three matters:

Guardrails: these are secure blocks that forestall a Massive Language Mannequin (LLM) from responding about some matters.

Agent Analysis: Have you ever ever considered how correct the responses from LLM are? I guess you probably did. So we are going to see the primary methods to measure that.

Monitoring: We may even study in regards to the built-in monitoring app in Agno’s framework.

We will start now.

Guardrails

Our first subject is the best, in my view. Guardrails are guidelines that can maintain an AI agent from responding to a given subject or listing of matters.

I imagine there’s a good likelihood that you’ve ever requested one thing to ChatGPT or Gemini and obtained a response like “I can’t speak about this subject”, or “Please seek the advice of knowledgeable specialist”, one thing like that. Normally, that happens with delicate matters like well being recommendation, psychological circumstances, or monetary recommendation.

These blocks are safeguards to stop individuals from hurting themselves, harming their well being, or their pockets. As we all know, LLMs are skilled on large quantities of textual content, ergo inheriting a variety of unhealthy content material with it, which might simply result in unhealthy recommendation in these areas for individuals. And I didn’t even point out hallucinations!

Take into consideration what number of tales there are of people that misplaced cash by following funding suggestions from on-line boards. Or how many individuals took the fallacious drugs as a result of they examine it on the web.

Properly, I assume you bought the purpose. We should forestall our brokers from speaking about sure matters or taking sure actions. For that, we are going to use guardrails.

The most effective framework I discovered to impose these blocks is Guardrails AI [1]. There, you will note a hub filled with predefined guidelines {that a} response should comply with with the intention to cross and be exhibited to the person.

To get began shortly, first go to this hyperlink [2] and get an API key. Then, set up the bundle. Subsequent, sort the guardrails setup command. It’ll ask you a few questions you can reply n (for No), and it’ll ask you to enter the API Key generated.

pip set up guardrails-ai
guardrails configure

As soon as that’s accomplished, go to the Guardrails AI Hub [3] and select one that you just want. Each guardrail has directions on how you can implement it. Mainly, you put in it by way of the command line after which use it like a module in Python.

For this instance, we’re selecting one known as Prohibit to Matter [4], which, as its title says, lets the person speak solely about what’s within the listing. So, return to the terminal and set up it utilizing the code under.

guardrails hub set up hub://tryolabs/restricttotopic

Subsequent, let’s open our Python script and import some modules.

# Imports
from agno.agent import Agent
from agno.fashions.google import Gemini
import os

# Import Guard and Validator
from guardrails import Guard
from guardrails.hub import RestrictToTopic

Subsequent, we create the guard. We are going to limit our agent to speak solely about sports activities or the climate. And we’re limiting it to speak about shares.

# Setup Guard
guard = Guard().use(
    RestrictToTopic(
        valid_topics=["sports", "weather"],
        invalid_topics=["stocks"],
        disable_classifier=True,
        disable_llm=False,
        on_fail="filter"
    )
)

Now we will run the agent and the guard.

# Create agent
agent = Agent(
    mannequin= Gemini(id="gemini-1.5-flash",
                  api_key = os.environ.get("GEMINI_API_KEY")),
    description= "An assistant agent",
    directions= ["Be sucint. Reply in maximum two sentences"],
    markdown= True
    )

# Run the agent
response = agent.run("What is the ticker image for Apple?").content material

# Run agent with validation
validation_step = guard.validate(response)

# Print validated response
if validation_step.validation_passed:
    print(response)
else:
    print("Validation Failed", validation_step.validation_summaries[0].failure_reason)

That is the response once we ask a couple of inventory image.

Validation Failed Invalid matters discovered: ['stocks']

If I ask a couple of subject that’s not on the valid_topics listing, I may even see a block.

"What is the primary soda drink?"
Validation Failed No legitimate subject was discovered.

Lastly, let’s ask about sports activities.

"Who's Michael Jordan?"
Michael Jordan is a former skilled basketball participant extensively thought-about one in every of 
the best of all time.  He gained six NBA championships with the Chicago Bulls.

And we noticed a response this time, as it’s a legitimate subject.

Let’s transfer on to the analysis of brokers now.

Agent Analysis

Since I began finding out LLMs and Agentic Ai, one in every of my essential questions has been about mannequin analysis. Not like conventional Information Science Modeling, the place you will have structured metrics which are enough for every case, for AI Brokers, that is extra blurry.

Thankfully, the developer neighborhood is fairly fast find options for nearly every part, and they also created this good bundle for LLMs analysis: deepeval.

DeepEval [5] is a library created by Assured AI that gathers many strategies to judge LLMs and AI Brokers. On this part, let’s study a few the primary strategies, simply so we will construct some instinct on the topic, and in addition as a result of the library is kind of in depth.

The primary analysis is essentially the most primary we will use, and it’s known as G-Eval. As AI instruments like ChatGPT change into extra widespread in on a regular basis duties, we’ve to ensure they’re giving useful and correct responses. That’s the place G-Eval from the DeepEval Python bundle is available in.

G-Eval is sort of a good reviewer that makes use of one other AI mannequin to judge how properly a chatbot or AI assistant is performing. For instance. My agent runs Gemini, and I’m utilizing OpenAI to evaluate it. This technique takes a extra superior strategy than a human one by asking an AI to “grade” one other AI’s solutions based mostly on issues like relevance, correctness, and readability.

It’s a pleasant solution to take a look at and enhance generative AI techniques in a extra scalable manner. Let’s shortly code an instance. We are going to import the modules, create a immediate, a easy chat agent, and ask it a couple of description of the climate for the month of Could in NYC.

# Imports
from agno.agent import Agent
from agno.fashions.google import Gemini
import os
# Analysis Modules
from deepeval.test_case import LLMTestCase, LLMTestCaseParams
from deepeval.metrics import GEval

# Immediate
immediate = "Describe the climate in NYC for Could"

# Create agent
agent = Agent(
    mannequin= Gemini(id="gemini-1.5-flash",
                  api_key = os.environ.get("GEMINI_API_KEY")),
    description= "An assistant agent",
    directions= ["Be sucint"],
    markdown= True,
    monitoring= True
    )

# Run agent
response = agent.run(immediate)

# Print response
print(response.content material)

It responds: “Delicate, with common highs within the 60s°F and lows within the 50s°F. Count on some rain“.

Good. Appears fairly good to me.

However how can we put a quantity on it and present a possible supervisor or consumer how our agent is doing?

Right here is how:

Create a take a look at case passing the immediate and the response to the LLMTestCase class.
Create a metric. We are going to use the tactic GEval and add a immediate for the mannequin to check it for coherence, after which I give it the that means of what coherence is to me.
Give the output as evaluation_params.
Run the measure technique and get the rating and cause from it.

# Check Case
test_case = LLMTestCase(enter=immediate, actual_output=response)

# Setup the Metric
coherence_metric = GEval(
    title="Coherence",
    standards="Coherence. The agent can reply the immediate and the response is smart.",
    evaluation_params=[LLMTestCaseParams.ACTUAL_OUTPUT]
)

# Run the metric
coherence_metric.measure(test_case)
print(coherence_metric.rating)
print(coherence_metric.cause)

The output seems to be like this.

0.9
The response immediately addresses the immediate about NYC climate in Could, 
maintains logical consistency, flows naturally, and makes use of clear language. 
Nevertheless, it might be barely extra detailed.

0.9 appears fairly good, on condition that the default threshold is 0.5.

If you wish to verify the logs, use this subsequent snippet.

# Examine the logs
print(coherence_metric.verbose_logs)

Right here’s the response.

Standards:
Coherence. The agent can reply the immediate and the response is smart.

Analysis Steps:
[
    "Assess whether the response directly addresses the prompt; if it aligns,
 it scores higher on coherence.",
    "Evaluate the logical flow of the response; responses that present ideas
 in a clear, organized manner rank better in coherence.",
    "Consider the relevance of examples or evidence provided; responses that 
include pertinent information enhance their coherence.",
    "Check for clarity and consistency in terminology; responses that maintain
 clear language without contradictions achieve a higher coherence rating."
]

Very good. Now allow us to study one other fascinating use case, which is the analysis of process completion for AI Brokers. Elaborating a little bit extra, how our agent is doing when it’s requested to carry out a process, and the way a lot of it the agent can ship.

First, we’re making a easy agent that may entry Wikipedia and summarize the subject of the question.

# Imports
from agno.agent import Agent
from agno.fashions.google import Gemini
from agno.instruments.wikipedia import WikipediaTools
import os
from deepeval.test_case import LLMTestCase, ToolCall
from deepeval.metrics import TaskCompletionMetric
from deepeval import consider

# Immediate
immediate = "Search wikipedia for 'Time collection evaluation' and summarize the three details"

# Create agent
agent = Agent(
    mannequin= Gemini(id="gemini-2.0-flash",
                  api_key = os.environ.get("GEMINI_API_KEY")),
    description= "You're a researcher specialised in looking out the wikipedia.",
    instruments= [WikipediaTools()],
    show_tool_calls= True,
    markdown= True,
    read_tool_call_history= True
    )

# Run agent
response = agent.run(immediate)

# Print response
print(response.content material)

The consequence seems to be superb. Let’s consider it utilizing the TaskCompletionMetric class.

# Create a Metric
metric = TaskCompletionMetric(
    threshold=0.7,
    mannequin="gpt-4o-mini",
    include_reason=True
)

# Check Case
test_case = LLMTestCase(
    enter=immediate,
    actual_output=response.content material,
    tools_called=[ToolCall(name="wikipedia")]
    )

# Consider
consider(test_cases=[test_case], metrics=[metric])

Output, together with the agent’s response.

======================================================================

Metrics Abstract

  - ✅ Job Completion (rating: 1.0, threshold: 0.7, strict: False, 
analysis mannequin: gpt-4o-mini, 
cause: The system efficiently looked for 'Time collection evaluation' 
on Wikipedia and supplied a transparent abstract of the three details, 
absolutely aligning with the person's purpose., error: None)

For take a look at case:

  - enter: Search wikipedia for 'Time collection evaluation' and summarize the three details
  - precise output: Listed here are the three details about Time collection evaluation based mostly on the
 Wikipedia search:

1.  **Definition:** A time collection is a sequence of information factors listed in time order,
 usually taken at successive, equally spaced closing dates.
2.  **Purposes:** Time collection evaluation is utilized in varied fields like statistics,
 sign processing, econometrics, climate forecasting, and extra, wherever temporal 
measurements are concerned.
3.  **Function:** Time collection evaluation includes strategies for extracting significant 
statistics and traits from time collection knowledge, and time collection forecasting 
makes use of fashions to foretell future values based mostly on previous observations.

  - anticipated output: None
  - context: None
  - retrieval context: None

======================================================================

General Metric Go Charges

Job Completion: 100.00% cross fee

======================================================================

✓ Assessments completed 🎉! Run 'deepeval login' to save lots of and analyze analysis outcomes
 on Assured AI.

Our agent handed the take a look at with honor: 100%!

You’ll be able to study way more in regards to the DeepEval library on this hyperlink [8].

Lastly, within the subsequent part, we are going to study the capabilities of Agno’s library for monitoring brokers.

Agent Monitoring

Like I advised you in my earlier submit [9], I selected Agno to study extra about Agentic AI. Simply to be clear, this isn’t a sponsored submit. It’s simply that I believe that is the best choice for these beginning their journey studying about this subject.

So, one of many cool issues we will reap the benefits of utilizing Agno’s framework is the app they make accessible for mannequin monitoring.

Take this agent that may search the web and write Instagram posts, for instance.

# Imports
import os
from agno.agent import Agent
from agno.fashions.google import Gemini
from agno.instruments.file import FileTools
from agno.instruments.googlesearch import GoogleSearchTools


# Matter
subject = "Wholesome Consuming"

# Create agent
agent = Agent(
    mannequin= Gemini(id="gemini-1.5-flash",
                  api_key = os.environ.get("GEMINI_API_KEY")),
                  description= f"""You're a social media marketer specialised in creating participating content material.
                  Search the web for 'trending matters about {subject}' and use them to create a submit.""",
                  instruments=[FileTools(save_files=True),
                         GoogleSearchTools()],
                  expected_output="""A brief submit for instagram and a immediate for an image associated to the content material of the submit.
                  Do not use emojis or particular characters within the submit. If you happen to discover an error within the character encoding, take away the character earlier than saving the file.
                  Use the template:
                  - Publish
                  - Immediate for the image
                  Save the submit to a file named 'submit.txt'.""",
                  show_tool_calls=True,
                  monitoring=True)

# Writing and saving a file
agent.print_response("""Write a brief submit for instagram with suggestions and methods that positions me as 
                     an authority in {subject}.""",
                     markdown=True)

To watch its efficiency, comply with these steps:

Go to https://app.agno.com/settings and get an API Key.
Open a terminal and sort ag setup.
If it’s the first time, it would ask for the API Key. Copy and Paste it within the terminal immediate.
You will notice the Dashboard tab open in your browser.
If you wish to monitor your agent, add the argument monitoring=True.
Run your agent.
Go to the Dashboard on the internet browser.
Click on on Classes. As it’s a single agent, you will note it beneath the tab Brokers on the highest portion of the web page.

Agno Dashboard after operating the agent. Picture by the creator.

The cools options we will see there are:

Data in regards to the mannequin
The response
Instruments used
Tokens consumed

That is the ensuing token consumption whereas saving the file. Picture by the creator.

Fairly neat, huh?

That is helpful for us to know the place the agent is spending roughly tokens, and the place it’s taking extra time to carry out a process, for instance.

Properly, let’s wrap up then.

Earlier than You Go

We’ve got realized lots on this second spherical. On this submit, we coated:

Guardrails for AI are important security measures and moral pointers applied to stop unintended dangerous outputs and guarantee accountable AI conduct.
Mannequin analysis, exemplified by GEval for broad evaluation and TaskCompletion with DeepEval for brokers output high quality, is essential for understanding AI capabilities and limitations.
Mannequin monitoring with Agno’s app, together with monitoring token utilization and response time, which is significant for managing prices, making certain efficiency, and figuring out potential points in deployed AI techniques.