Search returns pages. A knowledge base returns the answer.

The five-stage anatomy that every system shares — gbrain, Graphiti, and whatever you build.

You are about to look at gbrain, Graphiti, GraphRAG, pgvector, and a dozen blog posts that each sound like a different thing. They are not. They are all the same five-stage pipeline with different parts swapped. Learn the pipeline once and every tool becomes "oh, it's just X with a graph bolted to retrieval."

The one idea: A knowledge base is a pipeline that turns raw signals into cited answers. Plain search stops at stage 4 and hands you a list. A knowledge base runs stage 5 and hands you a synthesized answer plus what it doesn't know.

The five stages

1 · IngestCapture a signal (meeting, email, tweet) as a stored unit.

2 · StoreKeep it in a system-of-record. Files, Postgres, a graph.

3 · RetrievePull the few units relevant to a query.

4 · SynthesizeLLM writes a cited answer + flags gaps.

5 · Link / enrichWire new units to existing ones; consolidate over time.

Numbered 1–5, but in a live system 5 loops back to 2: every write enriches the store, so the next retrieval is smarter. That loop is what makes it a brain and not a folder.

Where the two tools you care about sit

Same pipeline, different organs. Notice that nothing here is exotic — they just made different bets at each stage.

Stage	gbrain	Your Graphiti setup
Ingest	Agents capture meetings, email, voice, ideas	/handoff writes session memory
Store	Git markdown (source of truth) + Postgres/PGLite	Neo4j graph
Retrieve	Hybrid: vector + BM25 + RRF + graph signals	Graph traversal + embeddings
Synthesize	Cited prose + explicit gap analysis	Returned to /prime as context
Link	Auto-wires typed edges on write, 0 LLM calls	Temporal edges with validity windows

You already run a knowledge base. Graphiti is stages 1–3 and 5 done well, with synthesis delegated to Claude. gbrain is the same shape but bets on git files as the store and adds a hard synthesis layer on top. Neither is magic. Both are this table.

Why "returns the answer" matters for Rejig: a client doesn't want ten matching call transcripts. They want "DuckyCutz's busiest branch is Phuket Town, and here's the source." That is stage 5 + 4. Selling the answer, not the search box, is the product.

Retrieval practice

Close the table above. Answer from memory — the point is the effort, not the score.

What does a knowledge base do that plain keyword search does not?

In gbrain, what is the system-of-record at the store stage?

Which stage loops back and makes future retrieval smarter?

Read this next (primary source)

★ Why Vector Search Alone Isn't Enough: Hybrid Retrieval for RAG (InfoQ). It makes stage 3 concrete and sets up Lesson 2. ~10 min.

See also the course glossary for any term in this lesson.

This is a conversation, not a handout. Ask your teacher (me, in Claude Code) anything: "show me where Graphiti does stage 5 in my actual repo", "what would gbrain's store look like for DuckyCutz", or "I don't buy that synthesis is the hard part." Push back. That's how the next lesson gets aimed right.

Start of course Next → Lesson 2: Why retrieval needs both meaning and exact match