learn-rag-with-phoebe / Leader session 6 of 6
Learn RAG with Phoebe · Leader track · Session 6 of 6

From POC to production and where RAG is heading

The most dangerous moment in a RAG project is the one that feels like success: the demo works, everyone claps, and someone says "ship it". A working demo and a trusted production system are separated by a real distance - evaluation, governance, monitoring - and this final session is the roadmap across it. You will place your own project on the map, name the gate it must pass next, meet the team roles that make it real, and finally see where RAG is going: from a fixed pipeline to a tool an agent decides to call.

🟠 Deciding Leaders: C-level · managers · curious minds No code, ever 45 min
0-3 · Welcome 3-20 · Concepts 20-40 · Exercises 40-45 · Q&A
Part 0

The demo is the start line, not the finish

Everything you have learned across five sessions converges here. You know what RAG is, how it is built, what it costs, how to judge it, and how to govern it. Now the question is operational: how do you move from "it worked in the demo" to "we trust it in production, and we can prove it"? The answer is a series of gates, each adding one thing the last one lacked. A team that ships the demo skips the gates and pays for it later. A team that names its current gate - and what it must pass next - is doing the actual work of productionizing AI. That naming is what you leave this course able to do.

Live - presented in session Self-study - read after class ★ Try it now prompt Sources covered at the end
★ What you walk out with today The four-stage POC-to-production roadmap and the trap of shipping the POC, the team roles that carry a RAG system, the honest line between when plain RAG is enough and when you need an agent, contextual retrieval as today's quality frontier, and a warm sense of where to go next.
Part 1 · covers the four gates to production

The POC-to-prod roadmap 9 min live

Four stages, each one adding what the previous lacked. You cannot skip a stage; you can only pretend to, and the pretending shows up in production. Everything from sessions a1 through a5 slots onto this strip.

STAGE 1 POC Does retrieval work at all? the demo STAGE 2 Evaluated We have a scorecard the Triad + golden set STAGE 3 Governed Access, refusal, citations the a5 controls STAGE 4 Production Monitored, fresh, cost-capped and still measured The trap: shipping Stage 1 as if it were Stage 4. A demo proves it CAN work, not that it's trustworthy.
🔍 Click to zoom - four gates, each adding what the last one lacked
LiveThe trap of shipping the POC3 min

A proof-of-concept answers exactly one question: can retrieval work at all for this use case? That is genuinely valuable - it kills bad ideas cheaply. But it is a single hand-picked demo, not evidence of reliability. The trap is treating "it worked when I tried it" as "it works".

  • A POC has no scorecard. You watched a few answers look good. You do not know the faithfulness rate, the refusal behavior, or how it does on the hard 10% of questions.
  • A POC has no governance. No access filtering, no audit log, no refusal policy. Fine for a demo, a liability in production.
  • A POC is unmonitored. When quality drifts - and it will, as documents change - nobody is watching.
The sentence that saves projects "This is a great POC - what does it need to reach the next gate?" Said in the meeting where everyone wants to ship, that one question is often the difference between a system you are proud of and an incident you are explaining.
LiveWhat each stage adds3 min

Read the roadmap as a stack of guarantees, each stage earning the right to the next.

StageAddsGate to pass
POCProof retrieval can workDoes a good answer come out at all?
EvaluatedA scorecard on a golden setDo faithfulness and recall clear our thresholds?
GovernedAccess, refusal, citations, auditCan it refuse, cite, and respect permissions?
ProductionMonitoring, freshness, cost capsWill we notice when it drifts, and can we afford it?

Notice how the whole course maps on: a1-a2 got you to a POC, a3 made it affordable, a4 made it evaluated, a5 made it governed. This session is about not stopping until production.

Self-studyThe team that carries a RAG system2 min read

Production RAG is not one engineer's side project. Four roles keep it honest, and as a leader your job is to make sure each one has an owner:

  • Data owner. Decides what goes in the index, keeps it fresh, and retires stale documents. Owns the answer to "is our knowledge current?"
  • Retrieval engineer. Owns the search step - chunking, embeddings, re-ranking. The person you call when context recall is low.
  • Evaluator. Owns the scorecard and the golden set. Reports faithfulness and recall over time, and proves whether a change actually helped.
  • Domain expert. The person who knows whether an answer is actually right in the real world - the source of truth for the golden set, and the final judge of quality.

These need not be four people, but they are four hats. A system missing the evaluator hat ships on vibes; one missing the domain-expert hat optimizes numbers that do not match reality.

Part 2 · covers RAG vs agents and the 2026 frontier

RAG vs agents, and where this is heading 8 min live

The frontier is moving fast, so here is the honest, durable version. The big shift: RAG is becoming something an agent decides to use, rather than a fixed pipeline that always runs. Retrieval turns into a tool the model can call when it judges it needs to.

Agent decides WHICH tool, and WHETHER to retrieve Retrieve tool (RAG) fetch & ground on docs Other tools search web · call an API Answer directly when no retrieval needed Question from a user Agentic RAG: retrieval becomes a tool the model calls when it judges it's needed - not a fixed step.
🔍 Click to zoom - RAG as one tool an agent can choose to call
LiveWhen plain RAG is enough vs when you need an agent3 min

Do not reach for an agent because it sounds advanced. The distinction is about the shape of the task.

  • Plain RAG is enough when the job is "answer this question from our documents". One retrieve, one generate, done. This covers the large majority of real use cases - support bots, policy assistants, research helpers - and it is simpler to build, cheaper to run, and easier to govern.
  • You need an agent when the task requires several steps, or a choice between actions: search our docs, then look something up externally, then decide whether to retrieve again. An agent can call retrieval as one of several tools and decide when to use it.
  • The plain-language test: if the task is "look it up and tell me", that is RAG. If it is "figure out what to do, using lookup as one of your moves", that is an agent that uses RAG.
The discipline this course keeps returning to Start with the simplest thing that works. Plain RAG first; an agent only when a single retrieve-and-answer genuinely cannot do the job. Complexity you did not need is the most common way these projects get expensive and fragile.
LiveContextual retrieval - today's quality frontier3 min

If you want one concrete upgrade to ask about, it is contextual retrieval. The problem it solves: when you chop a document into chunks, each chunk loses the context around it - a paragraph that says "the threshold was raised to 30" no longer knows which threshold or when. Contextual retrieval adds a short blurb of surrounding context to each chunk before indexing, so the retrieved passage carries its own meaning.

  • Why it matters to you: it meaningfully reduces retrieval failures - the wrong-passage problem from session a4 - which is the root of a lot of bad answers. Sessions a3 and a4 carry the specific numbers.
  • How to treat it: as the current best-practice frontier for quality, worth asking your team whether they have adopted. Not a silver bullet - it still needs the same evaluation and governance - but a real step up.
Real world

The chunk that forgot what it meant. A team (anonymized) kept seeing their assistant confuse two similarly worded policies. The chunks had been split so finely that neither carried enough context to tell them apart. Adding a one-line context header to each chunk before indexing - contextual retrieval - separated them cleanly, without touching the model at all.

Self-studyThe honest "re-verify the tooling" note, and where to go next2 min read

One honest caveat to close on: this space moves fast. The frameworks, the tool names, the exact best-practice recipe - all of it shifts on a timescale of months. Agentic RAG in particular is an active frontier, and the specific way a framework wires it up this quarter may look different next quarter.

  • What is durable is everything in this course: the retrieve-augment-generate idea, the three doors, the RAG Triad, the governance controls, the roadmap. These are concepts, and concepts age well.
  • What to re-verify is the tooling - which library, which feature, which vendor. Treat any specific tool claim as "true as of when we checked", and check again before you commit.

And that is the leader track. Six sessions ago, "AI that uses our data" was a phrase in a vendor pitch. Now you can define RAG in a breath, sort the three doors, judge an answer on the Triad, govern the risks, and map a project from demo to production. That is a genuinely rare literacy. If you want to feel how it is actually built - the embeddings, the retrieval, the live playground - the builder track starts at b1, and the natural next step for hands-on orchestration is LangChain ↗. Thank you for spending the hours. Go ask your team the good questions.

Exercise 1 of 2

Place your project on the roadmap ★ 12 min · pen and paper

The whole course pays off in this one honest placement. No code - just your real project and the four gates.

Name a real RAG project (or proposed one) in your organization. Write the four stages - POC · Evaluated · Governed · Production - as a line.

Place your project honestly. Most "live" systems are actually at POC or Evaluated wearing a production badge. Resist the flattering placement.

Name the next gate it must pass and the one thing that gate requires - a golden set, an access filter, a monitoring plan. Be specific: "we need a 50-example golden set owned by [name]" beats "improve evaluation".

Write the sentence you will say in your next project meeting: "We're at ___, the next gate is ___, and to pass it we need ___ by ___." That sentence is your entire course, applied.

★ Your roadmap partner (paste into any chat AI)Help me place a RAG project on a four-stage roadmap: POC (retrieval works at all), Evaluated (scorecard on a golden set), Governed (access, refusal, citations, audit), Production (monitored, fresh, cost-capped). Ask me questions to figure out which stage my project is really at, challenge me if I over-place it, then help me name the exact next gate and the one thing needed to pass it. Here's my project: [describe it]
Exercise 2 of 2

The RAG-vs-agent decision for your top use case ★ 8 min · decision drill

One last decision to make with a clear head, before someone sells you an agent you do not need.

Take your top use case and describe the task in one sentence, starting with a verb: "answer questions about...", "look up and summarize...", "decide and act on...".

Apply the test: is it "look it up and tell me" (plain RAG) or "figure out what to do, using lookup as one move" (an agent that uses RAG)? Most use cases are the former.

If you leaned agent, list the multiple steps or choices that justify it. If you cannot name at least two genuine steps or a real branch, it is plain RAG dressed up - go simpler.

Write your verdict and the reason. Then note what quality frontier you would ask about first - contextual retrieval is a strong default - and who on your team owns finding out.

The graduation move Every one of these exercises ends the same way: the simplest thing that works, named owners, and a specific next step. That is not a RAG skill - it is how you make any AI project real. Carry it out of this course.
Homework

After the leader track ◐ 30 min total

★ Questions to ask your data team this week
  1. Honestly, which roadmap stage is each of our RAG projects at - and are we treating any POC as if it were production?
  2. What is the next gate our flagship project must pass, and what single thing does it need to get there?
  3. Do our four roles - data owner, retrieval engineer, evaluator, domain expert - each have a named owner?
  4. For our top use case, is this really plain RAG, or does it genuinely need an agent that calls retrieval as one tool?
  5. Have we adopted contextual retrieval, and how do we keep verifying our tooling as this space keeps moving?
Source material

Official sources covered

The leader track teaches from published engineering guidance - no vendor marketing. This page covers:

Anthropic · Contextual RetrievalPart 2 · adding context to chunks before indexing · the current quality frontier
LangChain · agentic RAG (high-level)Part 2 · retrieval as a tool the model decides to call · re-verify specifics, they move fast
"Redesign the workflow" productionizing framing (ours)Part 1 · the four-gate roadmap and the trap of shipping the POC
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · The four stages of the POC-to-production roadmap, in order, are...

Each stage adds what the last lacked: proof it works, then a scorecard, then governance, then production monitoring and freshness. Skipping a gate shows up in production.

2 · When is a full agent warranted over plain RAG?

"Look it up and tell me" is plain RAG. "Figure out what to do, using lookup as one move" is an agent that calls retrieval as a tool. Start simple.

3 · Contextual retrieval improves quality by...

Chunking strips a passage of its surrounding context; contextual retrieval restores it before indexing, cutting the wrong-passage failures that cause bad answers.

Leader session 6 cheat sheet · pin this

The four gatesPOC → Evaluated → Governed → Production. Each adds what the last lacked; you can't skip one.
The POC trapA demo proves retrieval CAN work, not that it's trustworthy. Don't ship Stage 1 as Stage 4.
The team hatsData owner · retrieval engineer · evaluator · domain expert. Each needs a named owner.
Plain RAG vs agent"Look it up and tell me" = RAG. "Figure out what to do, using lookup" = agent that calls RAG.
Agentic RAGRetrieval becomes a tool the model decides when to call - not a fixed retrieve-then-generate step.
Contextual retrievalAdd surrounding context to each chunk before indexing; today's quality frontier for retrieval.
Re-verify the toolingConcepts are durable; frameworks and features move fast. Re-check specifics before you commit.
The graduation moveSimplest thing that works, named owners, a specific next step. Carry it to any AI project.