Skip to content
Is there an AI for this?

FDE ACADEMY · MODULE 01 · LESSON 02

Cut the work until one part is deployable

A workflow is not an AI problem. Some steps in it are. This lesson separates retrieval from extraction from drafting from judgement, and finds the smallest slice that can be finished and measured.

Effort
About ninety minutes with the record from lesson 01 in front of you.
Skills
rag · llm-evaluation

Free and self-paced. No certification is offered.

Before you startYou have a written workflow record from lesson 01, with volumes and at least one exception path.

01Four things a model can do, and one it cannot

Nearly every workflow step a model helps with is one of four jobs. Retrieval: find the passage that answers this. Extraction: pull these fields out of this document into a structure. Drafting: produce a first version a person will edit. Classification: put this into one of these buckets. Each has a different failure mode, a different evaluation method and a different amount of engineering behind it.

The fifth job — judgement with consequences — is not one of them. Deciding whether to accept a clause, approve a payment or dismiss an employee is a step where the model may prepare the material and a person decides. That is not a compliance nicety; it is the difference between a system you can defend and one you cannot.

  • Retrieval fails by returning the wrong passage confidently. Measure it with known questions and known sources.
  • Extraction fails per field. Measure it per field, not per document.
  • Drafting fails by being plausible. Measure it by how much editing the reader does.
  • Classification fails on the rare classes. Measure it on the rare classes.

02Separate the model’s job from the system’s job

Most of the work in a deployment is not model work. Getting files out of the shared drive, converting scans, keeping an index current, deciding who may see what, showing a person the source next to the answer — none of that is the model, and all of it is what makes the difference between a demo and something people use on a Tuesday.

When you decompose, write two columns. If the left column is short and the right column is long, you have understood the deployment correctly.

03The smallest slice that is still end to end

Choose one slice that goes all the way through: a real artefact enters, a real person gets a result, and you can tell whether it was right. A slice that stops before the person is a prototype, and prototypes do not produce the feedback that changes the design.

One collection, one team, one question type is a good first slice. Everything on this site is built so that slice can be widened later without being rebuilt.

04Do this

PRACTICAL TASK

Decompose the workflow into jobs and pick one slice

Turn the workflow record into a table of steps classified by job type, then choose and defend the single slice you would deploy first.

What you need

  • Your workflow record from lesson 01

Steps

  1. 01

    Take every step in your record and mark it retrieval, extraction, drafting, classification, judgement, or plumbing. Most will be plumbing. That is the correct result.

  2. 02

    For each model step, write the sentence you would use to check it: “given this input, the right answer is that.” If you cannot write the sentence, you cannot evaluate the step and it is not ready to be deployed.

  3. 03

    Read the deployment options section of the matching use-case page and note which option the site puts first for a neutral brief, and the caveat printed above it.

    Deployment options for internal search

  4. 04

    Now read a recipe end to end and count how many of its steps are model work. Open the flagship private knowledge base and look at its eleven deployment steps.

    Private company knowledge base — deployment steps

  5. 05

    Write one paragraph naming your first slice, the artefact that enters it, the person who receives the result, and the check that tells you it worked.

You are done when

  • Every step in your record carries a job label, and the plumbing column is longer than the model column.
  • Your chosen slice runs from a real artefact to a real person, and you can state its success check in one sentence.
  • You can say which of the eleven recipe steps are about the model, and which are about the system around it.

05Where these facts live

This lesson does not restate anything that is already recorded with its evidence elsewhere on the site. These are the pages it leans on.

Ticks are stored in your browser only. They are not sent anywhere, they are not attached to an account, and clearing your browser data removes them.