Building Agent Workflows with Gemini 2.5 Pro—Does It Hold Up?

Key Concepts:

• Gemini 2.5 Pro: A new language model with improved benchmarks, coding abilities, and a large context window.
• Function Calling: The ability of an LLM to use external tools or functions to answer user queries.
• Parallel Function Calls: Executing multiple function calls simultaneously to gather information from different sources.
• Text-to-SQL: Converting natural language queries into SQL queries to interact with databases.
• Agentic Systems: Systems where LLMs act as agents, using tools and functions to achieve specific goals.
• System Prompt: Instructions given to the LLM to guide its behavior and define its role.
• Chain of Thought: The internal reasoning process of the LLM, showing how it arrived at a particular answer.

1. Gemini 2.5 Pro Overview and Benchmarks

• Gemini 2.5 Pro is presented as a high-performing language model, exceeding previous models in benchmarks, particularly on the Live Bench.
• Independent benchmarks confirm the model's strong coding capabilities, as demonstrated by its performance on the Polyglot ADER benchmark.
• The model is available for free via AI Studio, with a knowledge cutoff date of January 2025.
• API limitations include a rate limit of 5 requests per minute or 50 requests per day.
• The model boasts a large context window, accepting up to 1 million input tokens and generating up to 65,000 output tokens.
• The API does not provide the chain of thought for responses, potentially to prevent model training on the reasoning process.

2. Basic Function Calling

• Function calling allows an LLM to interact with external tools to gather information or perform actions.
• The typical function calling flow involves the LLM determining if a tool is needed, selecting the appropriate tool, generating inputs for the tool, receiving a response, and generating a final response to the user.
• The Gemini SDK for Python automates function calling but also allows for manual control.
• A simple example is provided using a get_current_weather function that returns hardcoded weather data for specific locations (San Francisco, New York) or a default response for other locations.
• The function includes a docstring that describes its purpose and expected inputs.
• The model configuration includes a list of available tools, in this case, the get_current_weather function.
• The output shows the model's reasoning process, including the selection of the get_weather function and the input provided.

3. Parallel Function Calls

• The model demonstrates the ability to make parallel function calls, executing multiple functions simultaneously.
• An example is provided with two functions: get_current_weather and get_population.
• The prompt "compare the weather and population of New York and San Francisco" triggers the model to execute both functions in parallel.
• The model successfully retrieves weather and population data for both cities and provides a comparison.
• The speaker notes that the model has not struggled with parallel function calls, even with more than two functions.

4. Text-to-SQL Assistant

• The model is used to build a basic text-to-SQL assistant that converts natural language queries into SQL queries.
• A sample database is created using SQLite, containing employee data with columns for employee number, name, department, and salary.
• A function called execute_SQL_query executes SQL queries and returns the results.
• The function is passed as a tool to the model.
• A system prompt instructs the model to always check the database schema first.
• The query "what are the average salaries by department" is successfully translated into a SQL query, and the model returns the average salaries for each department.
• The intermediate SQL query generated by the model is shown.

5. Combining Function Calls: News and Sentiment Analysis

• The model combines multiple function calls to analyze news articles and determine sentiment.
• A get_company_news function retrieves news articles for different companies (Apple, Microsoft, Google).
• A sentiment_analyzer function analyzes the sentiment of a news article based on the presence of positive or negative words.
• The prompt "what is the recent news about Apple and what's the general sentiment" triggers the model to use both functions.
• The model retrieves news articles about Apple, analyzes the sentiment of each article, and provides an overall sentiment assessment.

6. Complex Function Calling: Trip Planning

• The model is used to plan a trip involving multiple steps and function calls.
• Functions include get_weather_information, search_flights, search_hotels, currency_conversion, and trip_planning.
• The prompt "I am planning a trip from New York to Paris from April 10 to 13 2025 I need help with finding flights from New York to Paris on the April 10th Checking the weather during my stay Recommending hotels for two people then planning for a daily attorney and here's my total budget" requires the model to use multiple functions sequentially and in parallel.
• The system prompt instructs the model to act as a travel planner assistant and use the available tools to gather information.
• The model generates a comprehensive trip plan, including flight options, weather information, hotel recommendations, a daily itinerary, and a budget breakdown.
• The model performs both sequential and parallel function calls to gather the necessary information.
• The speaker expresses a desire for the API to provide the internal chain of thought, showing how the model decided on the sequence of function calls.

7. Business Intelligence Dashboard: Combining Text-to-SQL and Unstructured Data

• The model is used to build a business intelligence dashboard that combines text-to-SQL with unstructured data analysis.
• The database includes multiple tables: sales, product, regions, and employees.
• Functions include execute_SQL_query, get_company_database_schema, get_market_data, analyze_sales_trends, and provide_competitive_analysis.
• The prompt "I need a comprehensive analysis of our electronics product category" requires the model to identify the relevant category, query the database, retrieve market data, and analyze sales trends.
• The system prompt instructs the model to act as a business intelligence assistant and follow a step-by-step process: query the database, retrieve external market data, and present a well-structured analysis.
• The model generates a comprehensive analysis of the electronics product category, including competitive analysis, market outlook, and sales trends.
• The model follows a step-by-step approach, querying the database, retrieving market data, and generating a final response based on the gathered information.
• The speaker recommends providing step-by-step instructions to the model to guide its behavior, rather than relying on its probabilistic nature to make decisions.

8. Conclusion

• Gemini 2.5 Pro is a powerful model with strong reasoning and coding capabilities, making it suitable for complex workflows.
• The model's ability to perform function calling, parallel function calls, and text-to-SQL enables the creation of sophisticated agentic systems.
• Crafting a well-defined system prompt with step-by-step instructions is crucial for guiding the model's behavior and ensuring accurate results.
• The speaker offers consulting and advising services for businesses looking to solve similar problems.