Turn the architecture into a Rejig offering.

Scope, stack, and price a knowledge-base product for a real client, using the decision framework you now own.

Goal of this lesson: walk one real client (DuckyCutz) from "what questions must it answer" to a defensible architecture, a build plan, and a price. This is tier-4 of your mission, and it reuses every prior lesson as a checklist.

Step 1 · Start from the questions, not the tech (L7 axis 1)

Never pitch "a knowledge base." Pitch the answers it gives. List the questions the client actually has:

DuckyCutz question	Shape	Implies
"What are Patong's hours?"	passage lookup	vector + BM25
"Which barber is good with kids?"	trait lookup, free text	vector (multilingual!)
"What's order DKC-0083?"	exact id	BM25
"Who's our most-rebooked barber this month?"	relational + temporal	graph, IF they need it

Three of four are flat. The fourth is the only one that wants a graph, and it may not be worth it yet. That table is your scoping conversation.

Step 2 · Apply the two-axis test (L7)

DuckyCutz's KB is mostly static (hours, services, barber traits change rarely) and mostly passage-shaped. Verdict from Lesson 7: vector RAG + BM25 + multilingual embeddings. No graph. Resisting the graph here is the senior move, you'd be selling complexity the client pays to maintain and never uses.

The ladder, applied to a sale (L1): the cheapest architecture that answers the client's real questions is the right one. Upsell the graph later, when "who's our most-rebooked barber" becomes a question they'll actually pay for. Scope the simple thing, ship, expand.

Step 3 · The stack (reuse what you run)

Ingestclient docs + call transcripts → chunks (L4)

StorePostgres + pgvector (Vercel/Supabase)

Retrievehybrid + RRF + rerank (L2, L3)

Synthesizegrounded, cited, gap-aware (L8)

Reachinto VAPI agent / GHL / web widget

You already operate this shape: the DuckyCutz VAPI agent has a KB, you run Postgres-adjacent infra daily, and you ship on Vercel. The product is not new tech, it is productizing the pipeline you understand into a repeatable client SKU.

Step 4 · Multilingual is the moat (L4)

Your clients are Thai. A KB that retrieves correctly across Thai, English, and Russian (DuckyCutz already serves all three) is a real differentiator most generic RAG vendors fumble. Lead with it. Pick a multilingual embedder, test Thai retrieval explicitly, and make "works in Thai" a checklist item, not an afterthought.

Step 5 · Price the answer, not the infra

You already have a pricing instinct from the usage portal (cost-plus markup + retainer). For a KB product:

Component	Model
Build / setup	One-time: ingest their docs, tune retrieval, verify in their language
Hosting + queries	Monthly retainer (covers infra + LLM cost + margin), like the usage portal's model
Expansion	Add the graph / temporal layer as a Phase 2 upsell when relational questions appear

The same Phase-1-audit → Phase-2-implementation motion that is already your business model. A KB is just a well-scoped Phase 2 deliverable with a clear architecture behind it.

Capstone task (the real one)

#	Do this with me
1	Pick a client (DuckyCutz or bobe). Write their 6-10 real KB questions. We'll classify each by shape + velocity.
2	Decide the architecture from the two-axis test. Defend why you did NOT add a graph (or why you did).
3	Draft the one-page proposal: questions answered, stack, multilingual guarantee, price. I'll pressure-test it like the /devil command.

Mission check. If you can do this capstone task unaided, you've hit all four tiers: you understand the space, you run Graphiti, you can run gbrain, and you can scope a client product. That was the definition of done in MISSION.md.

Primary source

★ agentic-rag-knowledge-graph (GitHub): a reference build combining pgvector (semantic) + Neo4j/Graphiti (temporal) in one agent, the exact hybrid you'd grow a client into for Phase 2. Read its README as a blueprint.

Ask your teacher. This is where I'm most useful. Try: "let's scope the DuckyCutz KB proposal now", "what would Phase 2 (the graph upsell) actually cost to build?", or "council this architecture before I pitch it." You finished the course, so now we build.

← Lesson 10: Run gbrain Course complete · revisit the syllabus