Katkat
Back

Design a Retrieval-Grounded Local Sports News Recommender

MicrosoftMachine Learning Engineer · HR Screen
MediumMachine Learning EngineerHR ScreenML System Design

Last updated

This question tests the ability to design an end-to-end RAG system for news, focusing on temporal data relevance, hybrid retrieval strategies, and mitigation of LLM hallucinations.

System Design: Local Sports News Recommender

Design a system that provides personalized, summarized sports news updates for a user's local city. The system must ingest various real-time sports feeds (RSS, social media, local news APIs), generate summaries, and present them in a mobile feed.

Requirements

  1. Data Ingestion: Handle high-velocity incoming sports articles while filtering for location-specific relevance.
  2. Retrieval Engine: Propose a hybrid retrieval approach that balances keyword-based search (for specific team names or player entities) and dense vector search (for semantic intent).
  3. Summarization: Use an LLM to synthesize multiple sources into a coherent local brief. Include provenance (linking back to sources).
  4. Hallucination Mitigation: Implement mechanisms to ensure the summary is strictly grounded in the retrieved documents.
  5. Personalization: Rank the news based on a user's stated team preferences and historical reading behavior.

Scale Assumptions

  • 10,000+ local news sources.
  • Peak traffic: 50,000 requests per minute during major sporting events.
  • Latency target: Under 500ms for retrieval; under 3 seconds for generated summaries.

Key Deliverables

  • Sketch the high-level architecture: ingestion, indexing, retrieval, and generation stages.
  • Define the data model for articles and user profiles.
  • Propose an evaluation strategy for both retrieval accuracy and generation quality (hallucinations).
  • Discuss how to handle 'stale' news during live games and how to degrade the system gracefully if the LLM inference service is overloaded.

Hints

Hint 1

Consider how to represent 'local' metadata in your vector embeddings to ensure geographical relevance during the retrieval phase.

Hint 2

For hallucination checks, investigate NLI (Natural Language Inference) models or self-consistency loops to compare the summary against the source snippets.

Hint 3

Think about using a multi-tiered caching strategy (e.g., pre-computed summaries for breaking news) to handle high-traffic spikes.

Follow-up questions

  • How would you handle a situation where two major news sources provide conflicting scores or facts about a game?
  • If you move from a batch-processing ingestion pipeline to a streaming architecture (e.g., using Kafka), what changes in your system design?
  • How would you implement 'freshness' as a primary signal in your ranking function without completely suppressing relevant older stories?