rag and retrieval strategies

Explicitly mentions RAG and vectorless RAG as components of agentic AI stack, positioning them as data grounding layer

Similar problem to RAG: selecting relevant context from a large pool. Here it's selecting relevant trajectory reasoning; in RAG it's selecting relevant documents. Both use attention signals.

MCP resources interface is a retrieval mechanism for structured data access. Unlike traditional RAG (vector similarity), MCP enables exact retrieval + semantic understanding of what data exists. This

Supacrawl is a specialized form of RAG for structured data: instead of embedding text documents, it creates a queryable database snapshot agents can explore. The pattern of 'preserving structured cont

references/ folders function as context retrieval mechanism—agents read 'on demand' rather than loading all context upfront, similar to retrieval-augmented generation pattern.

Brave Search MCP server is a form of real-time retrieval—extending Claude's knowledge beyond training cutoff, similar to RAG pattern but via standardized protocol.

Article mentions JSON RAG as one tool type accessible via MCP servers, showing MCP can integrate with retrieval systems. Not primary focus but mentioned as capability.

The agents are retrieval-augmented by actual API behavior (observing error paths, categorizing retryable errors). This is retrieval in the form of behavioral observation rather than document retrieval

MCP servers can implement retrieval logic (e.g., Postgres queries via MCP). Understanding MCP topology choices affects how you design retrieval systems connected to LLMs.

While RAG retrieves static knowledge, Memento-Skills actively updates skills—suggests retrieval systems need evolution mechanisms.

FairQE's dynamic calibration and physiotherapy's real-time pose estimation suggest retrieval of specialized context (gender variants, pose data) as part of agent context.