Objective

Design and implement the LLM layer responsible for: (1) analyzing user news queries into structured intents/entities/clarity scores, and (2) generating concise, engaging article summaries and other LLM-driven enrichments (e.g., headlines, short blurbs). The LLM layer should expose clean input/output contracts for downstream services and be easy to test and tune independently of ingestion, storage, or trending simulation components.

LLM Interaction layer

There will be two LLM interactions

1. Analyse user query
2. Summarise news data

Analyse user query

Analyses user news queries and extracts structured information to enable precise news retrieval. Converts free-text queries into a machine-readable format, capturing user intent, entities, and query clarity.

Input

• query (string): The user’s news-related question or request.

Output

{
  "intents": [
    {
      "type": "category | score | search | source | nearby | trending",
      "intent_weight": 0.0,
      "entities": ["list of people, organizations, locations, events"],
      "locations": [{"lat": 0.0, "long": 0.0}] // optional, only for 'nearby'
    }
  ],
  "clarity_score": 0.0
}