MeMind News Engine MVP

Boundary

memind-news-engine owns collection, event clustering, ranking, and briefing candidates.
Memind consumes ranked candidates and owns page generation and delivery.
memind_adm owns source configuration, ranking policy editing, review, run monitoring, and publish controls.
The legacy ops/ UI is not extended.

Current pipeline

Memind SearXNG query groups
  -> memind-news-engine HTTP API
  -> event clustering by canonical URL and title similarity
  -> deterministic rule score
  -> optional structured LLM assessment
  -> global score plus optional user-interest score
  -> breaking-news override
  -> category quotas and source diversity
  -> ranked candidates in the existing briefing prompt
  -> existing page generation, dedupe, image, and WeChat delivery guards

The first slice is deliberately deterministic. The model may provide the six semantic dimensions, but application code calculates the weighted score and owns final selection.

Score contract

LLM assessment values are 0-10:

{
  "importance": 8,
  "novelty": 9,
  "impact": 8,
  "credibility": 9,
  "relevance": 8,
  "actionability": 7,
  "category": "ai_tech",
  "reason": "A flagship model release changes capability and API economics."
}

Semantic score weights are 25% importance, 20% impact, 15% novelty, 15% credibility, 15% relevance, and 10% actionability. When the LLM assessment is present, global score is 35% rule score plus 65% semantic score. When a user-interest score is present, final score is 70% global plus 30% personal.

Breaking override is active when importance is at least 9, there are at least two distinct sources, and credibility is at least 8.

Runtime controls

MEMIND_NEWS_ENGINE_DAILY_LIMIT: selected event count, default 15.
MEMIND_NEWS_ENGINE_MINIMUM_SCORE: minimum final score, default 0 during the observation phase.
MEMIND_NEWS_ENGINE_URL: independent service base URL; ranking is skipped when unset.
MEMIND_NEWS_ENGINE_API_TOKEN: service-to-service bearer token.
MEMIND_NEWS_ENGINE_TIMEOUT_MS: request timeout, default 5000.

Next slice

The engine repository now owns persistence migrations and the ranking API. Next work is the bounded LLM scoring worker, collection scheduler, and memind_adm management UI.

2.3 KiB Raw Blame History