Agentic AI 102: Guardrails and Agent Evaluation

Within the first put up of this sequence (Agentic AI 101: Starting Your Journey Building AI Agents), we talked concerning the fundamentals of making AI Brokers and launched ideas like reasoning, reminiscence, and instruments.

After all, that first put up touched solely the floor of this new space of the information trade. There’s a lot extra that may be performed, and we’re going to be taught extra alongside the way in which on this sequence.

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

On this put up, we’ll cowl three matters:

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

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

Monitoring: We may even be taught concerning the built-in monitoring app in Agno’s framework.

We will start now.

Guardrails

Our first matter is the only, in my view. Guardrails are guidelines that may preserve an AI agent from responding to a given matter or record of matters.

I consider there’s a good likelihood that you’ve ever requested one thing to ChatGPT or Gemini and obtained a response like “I can’t discuss this matter”, 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 forestall individuals from hurting themselves, harming their well being, or their pockets. As we all know, LLMs are educated on huge quantities of textual content, ergo inheriting quite a lot 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 ideas from on-line boards. Or how many individuals took the flawed medication as a result of they examine it on the web.

Effectively, I suppose you bought the purpose. We should forestall our brokers from speaking about sure matters or taking sure actions. For that, we’ll use guardrails.

The very best framework I discovered to impose these blocks is Guardrails AI [1]. There, you will note a hub stuffed with predefined guidelines {that a} response should observe in an effort to move and be exhibited to the consumer.

To get began rapidly, first go to this hyperlink [2] and get an API key. Then, set up the bundle. Subsequent, sort the guardrails setup command. It is going to ask you a few questions that you may 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 tips on how to implement it. Principally, 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 Limit to Subject [4], which, as its identify says, lets the consumer discuss solely about what’s within the record. So, return to the terminal and set up it utilizing the code beneath.

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’ll prohibit our agent to speak solely about sports activities or the climate. And we’re proscribing 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 after we ask a few inventory image.

Validation Failed Invalid matters discovered: ['stocks']

If I ask a few matter that’s not on the valid_topics record, I may even see a block.

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

Lastly, let’s ask about sports activities.

"Who's Michael Jordan?"
Michael Jordan is a former skilled basketball participant broadly thought of certainly one 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 matter.

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

Agent Analysis

Since I began finding out LLMs and Agentic Ai, certainly one of my major questions has been about mannequin analysis. In contrast to conventional Knowledge Science Modeling, the place you’ve gotten structured metrics which might be ample for every case, for AI Brokers, that is extra blurry.

Fortuitously, the developer neighborhood is fairly fast find options for nearly the whole lot, 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 be taught 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 intensive.

The primary analysis is essentially the most fundamental we will use, and it’s known as G-Eval. As AI instruments like ChatGPT develop into extra widespread in on a regular basis duties, we now have 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 sensible reviewer that makes use of one other AI mannequin to judge how nicely a chatbot or AI assistant is performing. For instance. My agent runs Gemini, and I’m utilizing OpenAI to evaluate it. This methodology takes a extra superior method than a human one by asking an AI to “grade” one other AI’s solutions primarily based on issues like relevance, correctness, and readability.

It’s a pleasant approach to check and enhance generative AI techniques in a extra scalable manner. Let’s rapidly code an instance. We’ll import the modules, create a immediate, a easy chat agent, and ask it a few 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: “Gentle, with common highs within the 60s°F and lows within the 50s°F. Anticipate some rain“.

Good. Appears fairly good to me.

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

Right here is how:

Create a check case passing the immediate and the response to the LLMTestCase class.
Create a metric. We’ll use the strategy GEval and add a immediate for the mannequin to check it for coherence, after which I give it the which means of what coherence is to me.
Give the output as evaluation_params.
Run the measure methodology and get the rating and purpose from it.

# Take a look at Case
test_case = LLMTestCase(enter=immediate, actual_output=response)

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

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

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. 
Nonetheless, it could possibly be barely extra detailed.

0.9 appears fairly good, provided 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 sensible.

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 attention-grabbing use case, which is the analysis of activity completion for AI Brokers. Elaborating somewhat extra, how our agent is doing when it’s requested to carry out a activity, 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 sequence 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 end result seems to be excellent. Let’s consider it utilizing the TaskCompletionMetric class.

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

# Take a look at 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

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

For check case:

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

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

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

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

General Metric Cross Charges

Process Completion: 100.00% move fee

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

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

Our agent handed the check with honor: 100%!

You’ll be able to be taught far more concerning the DeepEval library on this hyperlink [8].

Lastly, within the subsequent part, we’ll be taught the capabilities of Agno’s library for monitoring brokers.

Agent Monitoring

Like I advised you in my earlier put up [9], I selected Agno to be taught extra about Agentic AI. Simply to be clear, this isn’t a sponsored put up. It’s simply that I feel that is the most suitable choice for these beginning their journey studying about this matter.

So, one of many cool issues we will reap the benefits of utilizing Agno’s framework is the app they make out there 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


# Subject
matter = "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 partaking content material.
                  Search the web for 'trending matters about {matter}' and use them to create a put up.""",
                  instruments=[FileTools(save_files=True),
                         GoogleSearchTools()],
                  expected_output="""A brief put up for instagram and a immediate for an image associated to the content material of the put up.
                  Do not use emojis or particular characters within the put up. When you discover an error within the character encoding, take away the character earlier than saving the file.
                  Use the template:
                  - Submit
                  - Immediate for the image
                  Save the put up to a file named 'put up.txt'.""",
                  show_tool_calls=True,
                  monitoring=True)

# Writing and saving a file
agent.print_response("""Write a brief put up for instagram with ideas and tips that positions me as 
                     an authority in {matter}.""",
                     markdown=True)

To observe its efficiency, observe 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 concerning 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 kind of tokens, and the place it’s taking extra time to carry out a activity, for instance.

Effectively, let’s wrap up then.

Earlier than You Go

Now we have realized rather a lot on this second spherical. On this put up, we coated:

Guardrails for AI are important security measures and moral pointers applied to forestall 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 important for managing prices, guaranteeing efficiency, and figuring out potential points in deployed AI techniques.