Agentic_RAG/README.md

Documentation Retrieval Agent

Overview

This agent is designed to answer questions related to documents you uploaded to Vertex AI RAG Engine. It utilizes Retrieval-Augmented Generation (RAG) with the Vertex AI RAG Engine to fetch relevant documentation snippets and code references, which are then synthesized by an LLM (Gemini) to provide informative answers with citations.

This diagram outlines the agent's workflow, designed to provide informed and context-aware responses. User queries are processed by agent development kit. The LLM determines if external knowledge (RAG corpus) is required. If so, the VertexAiRagRetrieval tool fetches relevant information from the configured Vertex RAG Engine corpus. The LLM then synthesizes this retrieved information with its internal knowledge to generate an accurate answer, including citations pointing back to the source documentation URLs.

Agent Details

Attribute	Details
Interaction Type	Conversational
Complexity	Intermediate
Agent Type	Single Agent
Components	Tools, RAG, Evaluation
Vertical	Horizontal

Agent Architecture

Key Features

• Retrieval-Augmented Generation (RAG): Leverages Vertex AI RAG Engine to fetch relevant documentation.
• Citation Support: Provides accurate citations for the retrieved content, formatted as URLs.
• Clear Instructions: Adheres to strict guidelines for providing factual answers and proper citations.

Setup and Installation Instructions

Prerequisites

• Google Cloud Account: You need a Google Cloud account.
• Python 3.9+: Ensure you have Python 3.9 or a later version installed.
• Poetry: Install Poetry by following the instructions on the official Poetry website: https://python-poetry.org/docs/
• Git: Ensure you have git installed.

Project Setup with Poetry

1. Clone the Repository:

   git clone https://github.com/google/adk-samples.git
   cd adk-samples/python/agents/RAG
   

2. Install Dependencies with Poetry:

   Note for Linux users: If you get an error related to keyring during the installation, you can disable it by running the following command:

   poetry config keyring.enabled false
   

   This is a one-time setup.

   poetry install
   

   This command reads the pyproject.toml file and installs all the necessary dependencies into a virtual environment managed by Poetry.

3. Activate the Poetry Shell:

   poetry env activate
   

   This activates the virtual environment, allowing you to run commands within the project's environment. Make sure the environment is active. If not, you can also activate it through

   source .venv/bin/activate 
   

4. Set up Environment Variables: Rename the file ".env.example" to ".env" Follow the steps in the file to set up the environment variables.

5. Setup Corpus: If you have an existing corpus in Vertex AI RAG Engine, please set corpus information in your .env file. For example: RAG_CORPUS='projects/123/locations/us-central1/ragCorpora/456'.

   If you don't have a corpus setup yet, please follow "How to upload my file to my RAG corpus" section. The prepare_corpus_and_data.py script will automatically create a corpus (if needed) and update the RAG_CORPUS variable in your .env file with the resource name of the created or retrieved corpus.

How to upload my file to my RAG corpus

The rag/shared_libraries/prepare_corpus_and_data.py script helps you set up a RAG corpus and upload an initial document. By default, it downloads Alphabet's 2024 10-K PDF and uploads it to a new corpus.

1. Authenticate with your Google Cloud account:

   gcloud auth application-default login
   

2. Set up environment variables in your .env file: Ensure your .env file (copied from .env.example) has the following variables set:

   GOOGLE_CLOUD_PROJECT=your-project-id
   GOOGLE_CLOUD_LOCATION=your-location  # e.g., us-central1
   

3. Configure and run the preparation script:

   • To use the default behavior (upload Alphabet's 10K PDF): Simply run the script:

     python rag/shared_libraries/prepare_corpus_and_data.py
     

     This will create a corpus named Alphabet_10K_2024_corpus (if it doesn't exist) and upload the PDF goog-10-k-2024.pdf downloaded from the URL specified in the script.

   • To upload a different PDF from a URL: a. Open the rag/shared_libraries/prepare_corpus_and_data.py file. b. Modify the following variables at the top of the script:

     # --- Please fill in your configurations ---
     # ... project and location are read from .env ...
     CORPUS_DISPLAY_NAME = "Your_Corpus_Name"  # Change as needed
     CORPUS_DESCRIPTION = "Description of your corpus" # Change as needed
     PDF_URL = "https://path/to/your/document.pdf"  # URL to YOUR PDF document
     PDF_FILENAME = "your_document.pdf"  # Name for the file in the corpus
     # --- Start of the script ---
     

     c. Run the script:

     python rag/shared_libraries/prepare_corpus_and_data.py
     

   • To upload a local PDF file: a. Open the rag/shared_libraries/prepare_corpus_and_data.py file. b. Modify the CORPUS_DISPLAY_NAME and CORPUS_DESCRIPTION variables as needed (see above). c. Modify the main() function at the bottom of the script to directly call upload_pdf_to_corpus with your local file details:

     def main():
       initialize_vertex_ai()
       corpus = create_or_get_corpus() # Uses CORPUS_DISPLAY_NAME & CORPUS_DESCRIPTION
     
       # Upload your local PDF to the corpus
       local_file_path = "/path/to/your/local/file.pdf" # Set the correct path
       display_name = "Your_File_Name.pdf" # Set the desired display name
       description = "Description of your file" # Set the description
     
       # Ensure the file exists before uploading
       if os.path.exists(local_file_path):
           upload_pdf_to_corpus(
               corpus_name=corpus.name,
               pdf_path=local_file_path,
               display_name=display_name,
               description=description
           )
       else:
           print(f"Error: Local file not found at {local_file_path}")
     
       # List all files in the corpus
       list_corpus_files(corpus_name=corpus.name)
     

     d. Run the script:

     python rag/shared_libraries/prepare_corpus_and_data.py
     

More details about managing data in Vertex RAG Engine can be found in the official documentation page.

Running the Agent

You can run the agent using the ADK command in your terminal. from the root project directory:

1. Run agent in CLI:

   adk run rag
   

2. Run agent with ADK Web UI:

   adk web
   

   Select the RAG from the dropdown

Example Interaction

Here's a quick example of how a user might interact with the agent:

Example 1: Document Information Retrieval

User: What are the key business segments mentioned in Alphabet's 2024 10-K report?

Agent: According to Alphabet's 2024 10-K report, the key business segments are:

1. Google Services (including Google Search, YouTube, Google Maps, Play Store)
2. Google Cloud (offering cloud computing services, data analytics, and AI solutions)
3. Other Bets (including Waymo for autonomous driving technology) [Source: goog-10-k-2024.pdf]

Evaluating the Agent

The evaluation can be run from the RAG directory using the pytest module:

poetry run pytest eval

Evaluation Process

The evaluation framework consists of three key components:

1. test_eval.py: The main test script that orchestrates the evaluation process. It uses the AgentEvaluator from Google ADK to run the agent against a test dataset and assess its performance based on predefined criteria.

2. conversation.test.json: Contains a sequence of test cases structured as a conversation. Each test case includes:

   • A user query (e.g., questions about Alphabet's 10-K report)
   • Expected tool usage (which tools the agent should call and with what parameters)
   • Reference answers (ideal responses the agent should provide)

3. test_config.json: Defines evaluation criteria and thresholds:

   • tool_trajectory_avg_score: Measures how well the agent uses the appropriate tools
   • response_match_score: Measures how closely the agent's responses match the reference answers

When you run the evaluation, the system:

1. Loads the test cases from conversation.test.json
2. Sends each query to the agent
3. Compares the agent's tool usage against expected tool usage
4. Compares the agent's responses against reference answers
5. Calculates scores based on the criteria in test_config.json

This evaluation helps ensure the agent correctly leverages the RAG capabilities to retrieve relevant information and generates accurate responses with proper citations.

Deploying the Agent

The Agent can be deployed to Vertex AI Agent Engine using the following commands:

python deployment/deploy.py

After deploying the agent, you'll be able to read the following INFO log message:

Deployed agent to Vertex AI Agent Engine successfully, resource name: projects/<PROJECT_NUMBER>/locations/us-central1/reasoningEngines/<AGENT_ENGINE_ID>

Please note your Agent Engine resource name and update .env file accordingly as this is crucial for testing the remote agent.

You may also modify the deployment script for your use cases.

Testing the deployed agent

After deploying the agent, follow these steps to test it:

1. Update Environment Variables:

   • Open your .env file.
   • The AGENT_ENGINE_ID should have been automatically updated by the deployment/deploy.py script when you deployed the agent. Verify that it is set correctly:

     AGENT_ENGINE_ID=projects/<PROJECT_NUMBER>/locations/us-central1/reasoningEngines/<AGENT_ENGINE_ID>
     

2. Grant RAG Corpus Access Permissions:

   • Ensure your .env file has the following variables set correctly:

     GOOGLE_CLOUD_PROJECT=your-project-id
     RAG_CORPUS=projects/<project-number>/locations/us-central1/ragCorpora/<corpus-id>
     

   • Run the permissions script:

     chmod +x deployment/grant_permissions.sh
     ./deployment/grant_permissions.sh
     

   This script will:

   • Read the environment variables from your .env file
   • Create a custom role with RAG Corpus query permissions
   • Grant the necessary permissions to the AI Platform Reasoning Engine Service Agent

3. Test the Remote Agent:

   • Run the test script:

     python deployment/run.py
     

   This script will:

   • Connect to your deployed agent
   • Send a series of test queries
   • Display the agent's responses with proper formatting

The test script includes example queries about Alphabet's 10-K report. You can modify the queries in deployment/run.py to test different aspects of your deployed agent.

Customization

Customize Agent

You can customize system instruction for the agent and add more tools to suit your need, for example, google search.

Customize Vertex RAG Engine

You can read more about official Vertex RAG Engine documentation for more details on customizing corpora and data.

Plug-in other retrieval sources

You can also integrate your preferred retrieval sources to enhance the agent's capabilities. For instance, you can seamlessly replace or augment the existing VertexAiRagRetrieval tool with a tool that utilizes Vertex AI Search or any other retrieval mechanism. This flexibility allows you to tailor the agent to your specific data sources and retrieval requirements.

Troubleshooting

Quota Exceeded Errors

When running the prepare_corpus_and_data.py script, you may encounter an error related to API quotas, such as:

Error uploading file ...: 429 Quota exceeded for aiplatform.googleapis.com/online_prediction_requests_per_base_model with base model: textembedding-gecko.

This is especially common for new Google Cloud projects that have lower default quotas.

Solution:

You will need to request a quota increase for the model you are using.

1. Navigate to the Quotas page in the Google Cloud Console: https://console.cloud.google.com/iam-admin/quotas
2. Follow the instructions in the official documentation to request a quota increase: https://cloud.google.com/vertex-ai/docs/quotas#request_a_quota_increase

Disclaimer

This agent sample is provided for illustrative purposes only and is not intended for production use. It serves as a basic example of an agent and a foundational starting point for individuals or teams to develop their own agents.

This sample has not been rigorously tested, may contain bugs or limitations, and does not include features or optimizations typically required for a production environment (e.g., robust error handling, security measures, scalability, performance considerations, comprehensive logging, or advanced configuration options).

Users are solely responsible for any further development, testing, security hardening, and deployment of agents based on this sample. We recommend thorough review, testing, and the implementation of appropriate safeguards before using any derived agent in a live or critical system.