Agentic Search

Agentic Search is an advanced AI-powered workflow built to design and implement intelligent search systems using Gemini 2.0 and a ReAct agent. It orchestrates complex searches by dynamically integrating multiple APIs and leveraging a flexible tool registry. This workflow enables the synthesis of diverse data sources to deliver detailed, cohesive, and scalable solutions tailored to intricate queries. For a practical guide to building Agentic Search from scratch, refer to the Medium article.

Key Features

• Gemini 2.0 Integration: Leverages Gemini 2.0's advanced natural language understanding and multimodal reasoning capabilities.
• ReAct Agent Framework: Implements the ReAct framework for iterative reasoning and decision-making in complex tasks.
• Dynamic Tool Registry: Seamlessly integrates tools, including Wikipedia search and Google Trends, to expand functionality.
• Multimodal Support: Handles both text and image-based inputs to enrich query responses.
• Streamlined Interface: Built with a clean and interactive UI using Streamlit for intuitive user interactions.

Architecture

flowchart LR
    %% Define node styles
    classDef userNode fill:#fff,stroke:#a8a8a8,stroke-width:2px
    classDef taskNode fill:#ffecec,stroke:#ffd6d6,stroke-width:2px
    classDef agentNode fill:#f0f5ff,stroke:#e1e8ff,stroke-width:2px
    classDef toolNode fill:#ffffff,stroke:#e6e6e6,stroke-width:2px
    classDef envNode fill:#f2fff2,stroke:#e1ffe1,stroke-width:2px
    classDef defaultNode fill:#ffffff,stroke:#e6e6e6,stroke-width:2px

    User((User)):::userNode

    subgraph Input["Agentic Search UI"]
        direction TB
        Task[Task]
        Outcome[Outcome]
    end

    subgraph Agent[" ReAct Agent "]
        direction LR
        subgraph Core[" "]
            LLM["Gemini (LLM)"]
        end
        
        subgraph Tools[" "]
            API["Tools (APIs)"]
        end
        
        subgraph Memory[" "]
            Mem[Memory]
        end
    end

    subgraph Actions[" Actions "]
        direction TB
        Act["fa:fa-gear Actions"]
    end

    subgraph Environments[" Environments "]
        direction TB
        
        subgraph GoogleServices[" Google Services "]
            direction LR
            Search["Search"]
            News["News"]
            Maps["Maps"]
            Images["Images"]
            Shopping["Shopping"]
            Finance["Finance"]
            Trends["Trends"]
            Events["Events"]
            Play["Play Store"]
        end

        subgraph Knowledge[" Knowledge "]
            direction LR
            Wiki["Wikipedia"]
            Trivia["Facts & Trivia"]
            Lyrics["Lyrics"]
        end

        subgraph Demographics[" Demographics "]
            direction LR
            Age["Age Prediction"]
            Gender["Gender"]
            Nationality["Nationality"]
        end

        subgraph Location[" Location Services "]
            direction LR
            Zip["ZIP Info"]
            IP["IP Info"]
            ISS["ISS Location"]
        end

        subgraph Media[" Media & Images "]
            direction LR
            Dogs["Dog Images"]
            Fox["Fox Images"]
            MultiModal["Multimodal"]
        end

        subgraph Commerce[" Commerce "]
            direction LR
            Walmart["Walmart"]
        end
    end

    %% Connections with enhanced arrows
    User ====> Task
    Task ====> LLM
    LLM ====> |"Reasoning Loop"| API
    API ====> LLM
    LLM ====> Mem
    Mem ====> LLM
    API <====> Act
    Act <====> |"observations"| Environments
    LLM ====> Outcome
    Outcome ====> User

    %% Apply styles
    class Task,Outcome taskNode
    class LLM,API,Mem agentNode
    class Search,News,Maps,Images,Shopping,Finance,Trends,Events,Play,Wiki,Trivia,Lyrics,Age,Gender,Nationality,Zip,IP,ISS,Dogs,Fox,MultiModal,Walmart envNode
    class Act defaultNode

    %% Container styles
    style Input fill:transparent,stroke:#ffd6d6,stroke-width:2px
    style Agent fill:transparent,stroke:#e1e8ff,stroke-width:2px
    style Actions fill:transparent,stroke:#e6e6e6,stroke-width:2px
    style Environments fill:transparent,stroke:#e1ffe1,stroke-width:2px
    style Core fill:transparent,stroke:#e1e8ff,stroke-width:2px
    style Tools fill:transparent,stroke:#e6e6e6,stroke-width:2px
    style Memory fill:transparent,stroke:#e6e6e6,stroke-width:2px

    %% Link styles - Updated with consistent pastel colors
    linkStyle 0,1 stroke:#ff9999,stroke-width:3px
    linkStyle 2,3,4,5 stroke:#99b3ff,stroke-width:3px
    linkStyle 6,7 stroke:#b3d9b3,stroke-width:3px
    linkStyle 8,9 stroke:#ff9999,stroke-width:3px

Loading

Prerequisites

1. Create a folder named credentials.
2. Inside the folder, create a .yml file containing API keys for Google and SerpAPI as shown below:

   GOOGLE_API_KEY: xxxxxxxxx
   SERP_API_KEY: xxxxxx

Setup

Clone the repository and install the required dependencies:

git clone https://github.com/arunpshankar/AgenticSearch.git
cd Agentic-Search

# Create a virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Upgrade pip and install dependencies
pip install --upgrade pip
pip install -r requirements.txt

Set the necessary environment variables:

export PYTHONDONTWRITEBYTECODE=1
export PYTHONPATH=$PYTHONPATH:.

Ensure your credentials folder contains the required API keys.

Usage

1. Run the Application:
   Launch the Agentic Search Streamlit app:

   streamlit run src/workflow/app.py

2. Provide Your Query:
   Enter your search query or upload an image via the interface.

3. View Results:
   Interact with the ReAct agent's reasoning trace and receive detailed, accurate answers.

Tools and APIs

Agentic Search integrates a variety of tools defined in registry.py, enabling diverse functionalities:

• Wikipedia Tools:

  • get_wiki_search_results: Fetch summaries and metadata from Wikipedia.

• Facts and Trivia:

  • get_cat_fact: Retrieve a random cat fact.
  • get_multiple_cat_facts: Fetch multiple cat facts.
  • get_cat_breeds: Retrieve a list of cat breeds.
  • get_random_joke: Fetch a random joke.
  • get_ten_random_jokes: Retrieve ten random jokes.
  • get_random_joke_by_type: Fetch a random joke of a specific type.
  • get_trivia_questions: Retrieve trivia questions.

• Animal Images:

  • get_random_dog_image: Fetch a random dog image.
  • get_multiple_dog_images: Retrieve multiple random dog images.
  • get_random_dog_breed_image: Fetch an image of a specific dog breed.
  • get_random_fox_image: Fetch a random fox image.

• Demographic Predictions:

  • get_predicted_age_by_name: Predict age based on a name.
  • get_gender_by_name: Predict gender based on a name.
  • get_nationality_by_name: Predict nationality based on a name.

• Location and Public Data:

  • get_zip_info: Retrieve location data for U.S. ZIP codes.
  • get_public_ip: Fetch the public IP address of the requester.
  • get_iss_location: Get the current location of the International Space Station.

• Google and SerpAPI Tools:

  • get_google_search_results: Perform a Google search.
  • get_google_image_search_results: Fetch Google Images search results.
  • get_google_news_search: Perform a Google News search.
  • get_google_maps_search: Search for places using Google Maps.
  • get_google_maps_place: Retrieve details of a specific place.
  • get_google_jobs_search: Perform a Google Jobs search.
  • get_google_shopping_search: Fetch Google Shopping results.
  • get_google_local_basic_search: Perform a local business search.
  • get_google_play_query_search: Search for apps in the Google Play Store.
  • get_google_events_basic_search: Retrieve event details from Google.
  • get_google_videos_basic_search: Perform a Google Videos search.
  • get_google_finance_basic_search: Fetch Google Finance data.
  • get_google_finance_currency_exchange: Retrieve exchange rate data for currency pairs.

• Third-Party APIs:

  • get_walmart_basic_search: Search for products on Walmart.
  • get_lyrics: Retrieve song lyrics.
  • get_google_trends_interest_over_time: Fetch Google Trends interest-over-time data.

• Multimodal Reasoning:

  • get_multimodal_reasoning: Perform reasoning based on both text and image inputs.

This comprehensive tool registry allows Agentic Search to address diverse and intricate queries effectively.

Hands-On Examples

1. Finding Current Location and Identifying Locations of Interest

2. Identifying Patterns and Cultural Significance of Mexican Talavera Tiles

3. Accurate Breed Identification Through Multimodal Analysis

Contribution

We welcome contributions! Fork this repository and submit a pull request with detailed descriptions of your updates.

License

This project is licensed under the MIT License. See the LICENSE file for details.