RAG IS A HACK. Chunking destroys meaning. I moved to Context Density. Large windows + Prompt Caching = Superior Reasoning.
WHY CONTEXT DENSITY?
- STOP CHUNKING: Feed entire vaults into the context. No fragments. No loss of structural meaning.
- GLOBAL REASONING: Ask complex timeline questions. RAG finds snippets; CAG sees the whole story.
- PROMPT CACHING: Instant queries. 90% cheaper. Performance is a feature, not an afterthought.
THE STACK: OBSIDIAN AGENTIC RAG
- CACHE: Upload vault state once. Let the model ‘know’ your world.
- NAVIGATE: Agents walk the graph, following links, not just keyword matches.
- SYNTHESIZE: High-density reasoning over full context.
Retrieval compared with context-first input
flowchart LR
accTitle: Retrieval versus context-first flow
accDescr: RAG searches an index for top-k chunks before answering. A context-first approach uses a vault snapshot and cached context alongside the question.
subgraph RAG["RAG: retrieve relevant chunks"]
direction LR
RQ["Question"] --> RI["Search index"]
RI --> RK["Top-k chunks"]
RK --> RP["Prompt"]
RP --> RA["Answer"]
end
subgraph CAG["CAG: reason over available context"]
direction LR
CV["Vault snapshot"] --> CC["Cached context"]
CQ["Question"] --> CR["Reasoning"]
CC --> CR
CR --> CA["Answer"]
end
A hand-drawn view of the same shift
Download the editable Excalidraw source
THE SHIFT
Vector DBs are overhead. Context windows are the new RAM. Navigate, don’t just search.