PowerUp Coaching — Living Knowledge Base

Insight: The Three-Layer Knowledge Architecture — Keyword, Graph, and Semantic Retrieval

5 min read

“You’re pulling relationships, similarity, and then the AI provides the relevance. You have a nice holistic approach to retrieval from your data.” — Lou

Session context: 2026-05-07_Mastermind — Lou elaborated on the knowledge architecture behind Kasimir’s Obsidian demo, naming the three retrieval layers that together produce what no single layer can achieve alone.

Core Idea

Every personal knowledge base has three ways information can be retrieved from it. Most practitioners use only one — usually keyword search — and wonder why their second brain feels shallow. The three layers are not alternatives to each other. They are complementary, each filling a gap the others leave.

Layer 1: Keyword search (grep/find). This is text-level retrieval — the system finds files containing specific words or phrases. Fast, exact, and excellent for known-term lookups. Its blind spot: it knows nothing about meaning or relationship. A document about “client retention” won’t surface when you search “churn reduction” unless those words appear literally in the file.

Layer 2: Knowledge graph (relationship traversal). This is the layer that keyword search cannot touch. When your vault defines entities (people, concepts, projects) and the relationships between them (produced, supports, references, contradicts), the AI can follow those connections. You can ask: “What did I think about X in the context of Y at time Z?” and the graph traversal finds the answer by walking relationships, not matching words. This also enables temporal search — impossible with semantic retrieval — because the relationships carry date metadata. A knowledge graph approaches RAG performance at roughly 96.6% for well-formed vaults, while doing things RAG cannot do at all.

Layer 3: Semantic retrieval (vector similarity). This is the layer most people associate with “AI search” — it finds content that is similar in meaning to your query, even if the words differ entirely. NotebookLM and Pinecone live here. But similarity is not the same as relevance. Semantic retrieval returns what is linguistically close to your question; it cannot tell you whether the result is the right document for your specific context, person, or time constraint.

The AI provides the fourth layer: relevance synthesis. Once the first three layers have surfaced candidates — by keyword, relationship, and similarity — Claude reasons across that raw material to produce the answer you actually need. The graph and taxonomy aren’t just about better search; they’re about giving the AI a better navigation map so its inference operates on the right context.

The practical implication of this architecture: taxonomy-first design matters more than the tools you choose. Kasimir’s system works because he defined his 5 content pillars before accumulating notes — every note can be categorized, weighted, and connected. Lou’s LKB works because it defines entity types, relationship classes, and temporal structure. Without taxonomy, even a well-built tool returns confusion. With taxonomy, even a modest Markdown vault approaches the performance of a vector database.

This also explains why building a knowledge base manually — tagging every file yourself — was always the real bottleneck. Claude Code can now do this automatically. The insight is knowing what to ask it to build: define the taxonomy first, then let AI classify incoming content against it. The taxonomy is the intelligence; the AI is the clerk.

Practical Application

The Three-Layer Architecture Audit

Before choosing or rebuilding a knowledge management tool, diagnose which layers you currently have:

  1. Keyword layer (Layer 1): Do you have a consistent naming convention and can you grep your vault reliably? If not, start here — folder structure and naming conventions before anything else.
  2. Graph layer (Layer 2): Are your entities named and are relationships explicit? Can you ask “what connects X to Y?” Do you have date metadata? This is where taxonomy investment pays the highest return.
  3. Semantic layer (Layer 3): Are you using NotebookLM, Pinecone, or similar for similarity search? This is the most overhyped layer — valuable, but incomplete without the other two.

The Taxonomy-First prompt: “I’m building a knowledge base for [your domain]. Here are the major entity types: [list]. Here are the relationship types that matter to me: [list]. Design me a folder structure, frontmatter schema, and linking convention that would support all three retrieval layers — keyword, graph, and semantic — from day one.”

Coaching Question:

“If you could ask your accumulated knowledge base five questions right now, what would they be? Which layer of retrieval would need to work to answer each one?”

Evolution Across Sessions

This establishes the baseline for the three-layer retrieval model as a named framework. Prior sessions introduced components: RAG design (Aug 14), multi-pass retrieval (Aug 7), local RAG vs. remote inference (Jul 17), and knowledge base as asset (Nov 6). This session names the three layers explicitly and adds the 96.6% performance claim for hybrid graph + markdown retrieval. Future sessions should test this against real vault sizes and document where the model breaks down.

Graph View

Backlinks