How do you implement RAG with real-time data sources?

Real-time RAG augments LLM responses with live, up-to-date information. Data sources: live APIs (weather, stock prices, news feeds), streaming databases (real-time analytics), web search (current events, recent publications), and internal systems (CRM records, support tickets, inventory levels). Architecture: implement a retrieval router that classifies whether the query needs real-time data, static knowledge base, or both. For real-time queries: translate the user question into appropriate API calls or search queries, execute retrieval in parallel with knowledge base search, merge real-time and static results with freshness weighting, and generate an answer with clear attribution of data recency. Implementation patterns: tool-augmented RAG (LLM decides when to call external tools for real-time data), scheduled refresh RAG (periodically update knowledge base from live sources), and hybrid RAG (combine cached knowledge with on-demand real-time retrieval). Web search integration: use search APIs (Tavily, Brave, Serper) for current information, extract and chunk search results, and include publication dates in context for temporal awareness. Challenges: real-time data may conflict with knowledge base (handle by preferring more recent sources for time-sensitive topics), API latency affects response time (implement timeouts and fallbacks), and rate limits constrain throughput. Caching: cache real-time results with short TTL (minutes for stock prices, hours for news), implement cache-aside pattern with freshness checks, and pre-fetch predictable queries during off-peak hours. Quality: validate real-time data before including in context, implement source reputation scoring, and flag when real-time and static sources disagree.