You are entirely correct. Google recently executed a massive philosophical and structural rebrand. Vertex AI has been officially replaced by the Gemini Enterprise Agent Platform, and the specific search component is now called Agent Search. [1, 2, 3, 4] This shift marks a move away from just managing “raw machine learning models” to managing “agentic systems”. The old Vertex AI capabilities (Model Garden, Pipelines, Custom Training) have been nested inside this new workspace. [2, 4, 5] If you are using Agent Search to power your social video platform’s recommendations, here is exactly how it fits into your Python architecture.

Instead of manually chunking content, generating your own vectors, and writing low-level K-Nearest Neighbors (KNN) algorithms, Agent Search acts as a fully managed retrieval-augmented generation (RAG) and recommendation API out of the box. [6, 7] For a Python social platform, it supports two distinct vertical-specific schemas: [7]

1. Media Recommendations: Out-of-the-box system tailored for videos, articles, and music.
2. Custom App Data Search: A flexible system that uses deep semantic search to surface related content. [7, 8, 9]

[Video Metadata / User Interactions] ➡️ [BigQuery Data Store] ➡️ [Agent Search Engine] ➡️ [Python Client API Client]

Because Agent Search integrates natively with your Google Cloud infrastructure, you don’t need complex Python ML pipelines. [6, 10, 11, 12]

Agent Search connects directly to document databases or BigQuery. [13, 14]

• Create a BigQuery dataset containing your video_metadata (Video ID, Title, Transcripts, Description, Creator, Tags).
• Create a secondary table for your user_events (User ID, Watch Time, Likes, Skips).

• Search for Gemini Enterprise Agent Platform (the platform will seamlessly redirect you if you type the old “Vertex AI” query).
• Navigate to the Agents section and create a new App.
• Select Search App (under the hood, this provisions Agent Search).
• Link your app to the BigQuery video data store you set up in Step 1. Agent Search will automatically handle parsing the metadata, extracting semantics, and establishing structural relationships. [5, 6, 8, 13, 14]

The legacy google-cloud-discoveryengine library or the unified google-genai SDK can be used in your backend to return recommendations securely. [16] To serve a “More Like This” or “Personalised For You” feed to a user, query the specific engine data store using Python:

from google.cloud import discoveryengine_v1beta as discoveryengine
def get_agent_search_recommendations(user_id, current_video_id):
    # Initialize the client (endpoints and core libraries remain unbreaking)
    client = discoveryengine.RecommendationServiceClient()
    
    # Configure the data store location path
    serving_config = client.serving_config_path(
        project="your-gcp-project-id",
        location="global",
        data_store="your-video-data-store-id",
        serving_config="default_config"
    )
    
    # Structure the user event context
    user_event = discoveryengine.UserEvent(
        event_type="detail-page-view", # Tells the engine what action triggers it
        user_pseudo_id=user_id,
        documents=[discoveryengine.DocumentInfo(id=current_video_id)]
    )
    
    # Request recommended videos
    request = discoveryengine.RecommendRequest(
        serving_config=serving_config,
        user_event=user_event,
        page_size=5, # Top 5 videos
    )
    
    response = client.recommend(request=request)
    
    # Parse out recommended video documents
    recommended_videos = []
    for result in response.results:
        recommended_videos.append(result.document.id)
        
    return recommended_videos

• No Cold Starts: If you upload a brand new video, Agent Search uses Google’s background foundation models to immediately understand the context of the title and video text, matching it to similar creators before anyone has watched it. [10, 17]
• Hybrid Search Ability: It blends traditional keyword matching (e.g., matching text tags like #comedy) with deep semantic neural search natively. [8, 14]
• Zero Infrastructure Management: You do not need to host or scale a vector database (like Pinecone or Milvus) or worry about scaling embedding computation power as your active user base grows. [6, 18]

Would you like to explore how to format your BigQuery tables to seamlessly match the structural requirements of Agent Search, or look at how to pass live watch metrics using Python?

[1] https://docs.cloud.google.com [2] https://www.beri.net [3] https://dev.to [4] https://cloud.google.com [5] https://gcpstudyhub.com [6] https://codelabs.developers.google.com [7] https://docs.cloud.google.com [8] https://docs.langchain.com [9] https://partner.skills.google [10] https://docs.cloud.google.com [11] https://www.cleveroad.com [12] https://medium.com [13] https://www.youtube.com [14] https://docs.cloud.google.com [15] https://docs.cloud.google.com [16] https://github.com [17] https://cloud.google.com [18] https://oneuptime.com