Architecture
A retrieval-augmented pipeline: every item is embedded and stored, then each assessment draws on the most relevant memory across the entire history, not just the latest headlines.
Data sources
13 subreddits
Geopolitics, regional, and country-specific communities.
11 feeds
Major news outlets across South Asia.
Configurable
Global news aggregation, optional.
Processing pipeline
Collect all news from every source without filtering.
Convert text to vectors with the local MiniLM-L6-v2 model.
Persist embeddings in ChromaDB as durable memory.
Query semantically relevant items via vector search.
Generate the assessment with Gemini over retrieved context.
Vector memory
Limitations
Data usage policy
All data is collected from public sources and used solely for analysis, in compliance with each platform's API terms. Embeddings are stored locally in ChromaDB. Individual items are never displayed; only the aggregated analysis is presented.