Building a Secure Multi-Tool AI Agent with Riza and Google Gemini in Colab

Secure API Key Management in Google Colab

This tutorial starts by demonstrating how to securely manage API keys necessary for integrating Riza and Google Gemini in Google Colab. It uses multiple methods such as Colab secrets, environment variables, and hidden user input prompts to ensure credentials are handled safely and seamlessly.

Installing and Importing Essential Libraries

The necessary libraries including LangChain community tools, Google Gemini integration, Riza’s secure Python execution package, and dotenv support are installed and imported quietly in Colab. Standard Python modules for system operations, typing, date-time, and JSON processing are also included.

Defining Secure API Key Setup

A function setup_api_keys() is defined to retrieve API keys from various sources in a prioritized manner. If keys cannot be found in Colab secrets or environment variables, it prompts the user securely via hidden inputs. Instructions are provided if key setup fails.

Importing LangChain and Riza Components

Key components such as Riza's ExecPython tool, LangChain's Google Gemini wrapper, agent executor, prompt templates, memory buffers, and callback handlers are imported to build a multi-tool AI agent with rich capabilities.

Enhanced Callback Handler for Logging

An AdvancedCallbackHandler class is implemented to capture detailed logs of each tool invocation with timestamps and maintain metrics like total actions. This facilitates real-time monitoring and debugging of agent operations.

Custom Tools: Math and Text Analysis

Two specialized tools are created:

• MathTool safely evaluates complex mathematical expressions using restricted environments to prevent unsafe operations.
• TextAnalyzer performs comprehensive text analysis including character frequency, word count, average word length, and specific character counts, returning the results as formatted JSON.

Validating API Keys

A validation function ensures that both Google Gemini and Riza API keys are functional by executing test calls. The notebook exits gracefully if validation fails, prompting the user to check credentials.

Initializing Tools and Gemini Model

The Riza ExecPython tool is instantiated for secure Python execution. The MathTool and TextAnalyzer are wrapped as LangChain Tools. The Gemini generative model is initialized with custom parameters for temperature, token limits, and sampling strategies, with a fallback to a lighter model if needed.

Constructing the AI Agent

A structured prompt template guides the AI agent’s reasoning and tool usage. Conversation memory is set up to retain recent exchanges. The advanced callback handler manages logging. These are combined to create a tool-calling agent wrapped in an executor that controls execution flow, verbosity, and iteration limits.

Querying the Agent with Sample Questions

A helper function ask_question() sends queries to the agent, prints responses and execution summaries, and handles exceptions gracefully. Sample questions include counting letters, calculating compound interest, analyzing word frequency, and generating Fibonacci sequences.

Summary of the Agent's Capabilities

The integration of Riza's sandboxed Python execution with Google Gemini's generative AI, orchestrated by LangChain, produces a modular, secure, and powerful multi-tool AI agent. This framework supports complex computations, detailed text analytics, and stepwise reasoning, suitable for data processing, research prototyping, and educational demonstrations in a secure environment.