“How’s the weather today in Delhi?”
Simple question — but what if we wanted a conversational AI that could answer it, explain the temperature trend, and log every detail of its interaction for analysis?
Agentic systems are booming. But building a reliable production-ready AI agent involves more than just connecting an LLM. You need tool usage, observability, error handling, and real-time responses. In this guide, we’ll build a Gemini AI-powered weather assistant that does all this, with observability powered by Maxim and real-time data fetched via Tavily.
The Problem
Conversational agents are great, until something breaks and you don’t know why. Whether it’s tool errors, vague responses, or inflated costs, most agent pipelines lack end-to-end monitoring.
That’s where Maxim comes in, an observability SDK for LLM-based apps. And with Google Gemini’s new APIs, you can now create agents that reason well and support tool usage. Let’s bring these together and solve a common use case: a weather bot.
What We’ll Build
We’ll implement a Gemini-powered conversational agent that:
Responds to user messages in a natural chat format
Uses Tavily to fetch current weather for any location
Logs all traces, latency, costs, and tool usage into the Maxim Dashboard
Offers both manual and automatic trace management
We have also published a video explanation of this tutorial here -
Prerequisites
Make sure you’ve got:
Python 3.8 or above
API keys for:
Basic comfort with Python and REST APIs
Step 1: Install Dependencies
Install all required libraries:
pip install google-generativeai
pip install maxim-py
pip install tavily-python
pip install python-dotenv
Step 2: Configure Environment
Create and load environment variables for API keys:
import os
from dotenv import load_dotenv
load_dotenv()
os.environ["GEMINI_API_KEY"] = "your_gemini_api_key"
os.environ["MAXIM_API_KEY"] = "your_maxim_api_key"
os.environ["TAVILY_API_KEY"] = "your_tavily_api_key"
os.environ["MAXIM_REPO_ID"] = "your_maxim_repo_id"
These keys are loaded and set as environment variables so they can be accessed throughout your app without hardcoding them again.
Add validation to avoid misconfigurations:
required_vars = ["MAXIM_API_KEY", "MAXIM_REPO_ID", "GEMINI_API_KEY", "TAVILY_API_KEY"]
for var in required_vars:
if not os.getenv(var):
print(f"❌ {var} is missing")
else:
print(f"✅ {var} loaded")
This ensures that none of the required API keys are missing. It prevents runtime errors due to misconfiguration.
Step 3: Set Up Maxim Logging
from maxim import Maxim, LoggerConfig
logger = Maxim().logger()
This initializes the Maxim logger for your specific repository ID. From here on, all traces will be pushed to this log repo for tracking agent activity, latency, tool usage, and more.
Step 4: Configure Gemini Client (with Maxim)
from google import genai
from maxim.logger.gemini import MaximGeminiClient
gemini_client = MaximGeminiClient(
client=genai.Client(api_key=os.getenv("GEMINI_API_KEY")),
logger=logger
)
This wraps the Gemini client with Maxim’s logging layer. Now all requests/responses and tool calls are logged automatically.
Step 5: Set Up Tavily for Weather Retrieval
from tavily import TavilyClient
tavily_client = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
Tavily will handle real-time weather searches. It searches the internet and returns snippets from trusted sources.
🌦️ Step 6: Define Weather Fetch Function
def get_weather_with_tavily(location: str) -> str:
try:
query = f"current weather in {location} today temperature conditions"
results = tavily_client.search(query=query, search_depth="basic", max_results=3)
weather_info = [
r.get('content', '')[:200] + "..."
for r in results.get('results', [])
if any(word in r.get('content', '').lower() for word in ['temperature', 'weather', 'degrees', '°'])
]
return f"Weather in {location}:\\n" + "\\n".join(weather_info[:2]) if weather_info else "No weather info found."
except Exception as e:
return f"Error: Could not retrieve weather for {location} — {str(e)}"
This function:
Creates a search query using the location
Queries Tavily for relevant web content
Filters for useful weather details (temp, degrees, etc.)
Returns a clean response or a fallback message
Step 7: Create the Conversational Agent
This initializes the agent with:
Gemini client for conversation
Maxim logger for observability
Weather function for tool use
Conversation history to track context
from uuid import uuid4
from maxim.logger import TraceConfig
class ConversationalAgent:
def __init__(self, gemini_client, logger, weather_function):
self.gemini_client = gemini_client
self.logger = logger
self.weather_function = weather_function
self.history = []
def chat(self, message: str, use_external_trace: bool = False):
print(f"User: {message}")
self.history.append(f"User: {message}")
try:
config = {
"tools": [self.weather_function],
"system_instruction": "You're a helpful weather assistant. Use the weather tool when needed.",
"temperature": 0.7,
}
if use_external_trace:
trace = self.logger.trace(TraceConfig(id=str(uuid4())))
response = self.gemini_client.models.generate_content(
model="gemini-2.0-flash",
contents=message,
config=config,
trace_id=trace.id
)
trace.end()
else:
response = self.gemini_client.models.generate_content(
model="gemini-2.0-flash",
contents=message,
config=config,
)
self.history.append(f"AI: {response.text}")
print(f"AI: {response.text}")
return response.text
except Exception as e:
error = f"❌ Error: {str(e)}"
print(error)
return error
def get_history(self):
return "\\n".join(self.history)
def clear_history(self):
self.history = []
print("History cleared.")
Step 8: Run a Test
Initializes the weather agent and tests it with a prompt. You’ll see a response from Gemini and logs in Maxim.
agent = ConversationalAgent(
gemini_client=gemini_client,
logger=logger,
weather_function=get_weather_with_tavily
)
agent.chat("What's the weather in London today?")
Step 9: Launch Interactive Chat Mode
This creates a CLI-based chat experience with basic commands for exiting, clearing, and reviewing the conversation.
def interactive_chat():
print("Chat started. Type 'exit' to stop.")
while True:
msg = input("\\nUser: ").strip()
if msg.lower() in ["exit", "quit", "bye"]:
print("👋 Chat ended.")
break
elif msg.lower() == "history":
print(agent.get_history())
elif msg.lower() == "clear":
agent.clear_history()
else:
agent.chat(msg)
interactive_chat()
Step 10: Verify in Maxim Dashboard
Once running:
Head to your Maxim dashboard
View the repository logs
Drill into trace IDs to view:
Tokens used
Response latency
Tool call usage
Full chat history
Advanced Options
Trace Modes
Automatic (default): Easier setup, no trace ID management
Manual: Add TraceConfig for full control (e.g., for long chains or external tracing)
Model Config
{
"temperature": 0.7,
"max_output_tokens": 1000,
"tools": [weather_function],
"system_instruction": "You are a weather assistant..."
}
Final Thoughts
This tutorial shows how to build a real-world conversational agent, not just with AI, but with the operational maturity needed for production use.
✅ Gemini AI
✅ Tavily tool integration
✅ Maxim observability
By logging every interaction, response, tool call, and cost, you gain deep insight into how your agents behave and scale.
Try It Yourself
Looking to plug in your own tools or run multi-step workflows?
This architecture is modular, just replace the weather_function with another tool like search, stock lookup, or even a retrieval-augmented QA pipeline.
Happy building! ☀️🌧️⚡
Originally published at https://www.getmaxim.ai on June 13, 2025.
Top comments (0)