Design Grounded Local Sports Recommendations
Recommend timely local sports coverage with retrieval-grounded summaries, safe fallbacks, and measurable model cost and quality.
Problem statement
For this exercise, assume a sports app receives licensed local game updates and short articles from versioned feeds. A user chooses a region and may follow teams. The home feed recommends recent, relevant items with a short generated reason or summary.
Every factual claim in generated text, including a score or game status, must be supported by a current permitted source version; if the evidence is missing, stale, contradictory, or retracted, the service shows a plain linked card or abstains instead of inventing a fact.
A user can report an irrelevant or incorrect recommendation. The source report asks for a local-sports recommendation system with LLM/RAG and explicitly probes retrieval platform, fine-tuning versus prompting, latency optimization, and hallucination controls.
Design source ingestion and retraction, regional eligibility, candidate retrieval and ranking, evidence selection, answer generation, validation, serving, and feedback.
Compare managed vector search such as Azure AI Search with a self-hosted vector index such as Milvus for this bounded workload; compare prompting with fine-tuning for grounded summaries; and explain when quantization or distillation could reduce serving cost without violating quality gates.
Include evaluation and a red-team review path for unsupported claims. The numbers below are exercise assumptions, not reported interview constraints. Rights negotiation, live video, betting, ticket sales, and building a foundation model are outside scope.
What to build
- Ingest versioned regional sports articles and game updates, and honor updates or retractions before their evidence can support new generated claims.
- Return a region-eligible ranked feed with source-linked items and generated text only when current evidence supports each factual claim.
- When retrieval, generation, or validation cannot meet its budget, return a clearly labeled source-linked card or abstain rather than fabricate a score or event.
- Capture attributable impressions and correction reports, evaluate relevance and factual support by region, and gate model or retrieval changes before rollout.
At scale
- For this exercise, assume up to 2 million active short articles and game updates across regions, 50,000 accepted source versions per day, and a peak of 100 feed requests per second.
- For this exercise, target p95 feed response within 1.5 seconds; an evidence-backed plain-card fallback may be faster and need not wait for generation.
- For this exercise, an accepted update or retraction should change eligibility for new feed requests within two minutes; a stale or unverified score must never be emitted as a fresh generated fact.
- For this exercise, retain source-version and decision evidence for 30 days so sampled claims and user corrections can be audited.