rag retrieval strategy

Persistent knowledge graphs are an evolution of RAG, addressing its statelessness limitation. This is a direct extension improving retrieval efficiency.

RAG results are explicitly mentioned as one component of context engineering that must be balanced against other information sources.

Semantic memory retrieval pattern is functionally equivalent to RAG: external knowledge stored separately, retrieved based on query/task, injected into prompt.

Mentions dynamic context retrieval using 'user query and metadata' to identify what to fetch—aligns with RAG pattern but applied within agent context management.

The Intercom attack vector shows RAG failure mode: customer support agent retrieves data based on instructions embedded in retrieved context. This is a retrieval poisoning problem.

RAG retrieval is listed as one of the components managed within the context lifecycle in context engineering

MCP servers can implement RAG-like retrieval patterns. The protocol standardizes how retrieval systems expose context to LLMs.

LlamaIndex is a primary RAG framework; context engineering relates directly to what context gets retrieved and how it's prioritized.

MCP can be viewed as a protocol layer above RAG—standardizing how agents retrieve and request context from knowledge sources.

MCP provides architectural foundation for how retrieval systems expose relevant context to AI—moves beyond prompt-level RAG to protocol-level context exposure

Article explicitly mentions 'retrieve the most relevant' information, identifying retrieval as critical component of context engineering architecture

The partitioning + tiered memory architecture is a retrieval strategy: organize data so that relevant subset can be quickly surfaced. This is architecturally similar to RAG chunking and indexing strat

Just-in-time data fetching vs upfront metadata dumps is fundamentally a retrieval strategy question—when and how to fetch context for agent use.

MCP servers effectively implement RAG infrastructure at the protocol level, giving models standardized way to retrieve external knowledge without custom SDK integration.

Cron jobs updating documentation/skill files function similarly to RAG refresh cycles—ensuring agents have fresh, relevant context to retrieve from.

RAG decisions (what knowledge to retrieve and include) now have an additional constraint: competitive sensitivity. Should affect what goes in retrieval index.

The folder organization acts as a structured retrieval system—when you need context about 'skills', you know exactly where to look, similar to RAG document organization

The OpenCV vs Deep Learning example shows a scenario where external domain knowledge (OpenCV context) should have been retrieved/provided but wasn't, leading to false equivalence.

RAG is about structured information retrieval; DDD/SOA is about structured information boundaries. Both solve similar problem: how does one part of system access information from another without losin

Memory files are retrieved and used as context in future sessions—a RAG pattern where user's own work becomes the retrieval corpus.

The underlying failure suggests code review tools need better retrieval of relevant context (similar code patterns, related PRs, design decisions) rather than just diff injection.

Automated context assembly from metadata layer resembles RAG approach of retrieving relevant context before inference