Fawzooz
عربيCollaborate
Your journey
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
0 of 34 units

You are in Degree 3 · Practical masteryunit 4 of 6Ahead of you: A real piece of professional work, and the page that shows you checked it.

Degree 3 · Unit 3.4

Build your knowledge assistant

A general system knows the world, and it does not know you. It does not know your contracts, your meeting minutes, or the way you write. The answer is not to train a model of your own, which is expensive and which you do not need. The answer is retrieval-augmented generation: giving the system your knowledge at the moment of the question rather than at the moment of training.

Imagine an intelligent human assistant who has just joined your organisation. You do not send them off to university. You give them your archive and say: "when someone asks you something, search in here first, then answer, and tell them which file you took it from." That is RAG, almost literally.

FIG. 16 — How the knowledge assistant works
1
Chunking
Your documents are divided into short, coherent passages
2
Indexing
Each passage is turned into a vector and stored in a semantic index
3
Retrieval
When you ask, the passages closest in meaning — not in wording — are fetched
4
Generation
The system answers from those passages alone, and states its source

The single-sitting project

You do not need a programmer for this. Most platforms today let you create a "knowledge space" and upload your own files into it. The quality comes from the decisions you make, not from the platform you chose:

  1. 1Choose the sources firmly: twenty trustworthy documents beat two hundred containing cancelled drafts. A poor source produces a poor answer with high confidence.
  2. 2Clean before uploading: delete duplicates and superseded material, and name the files meaningfully and with dates.
  3. 3Write a governing instruction: "Answer from the attached sources only, and name the file. If you do not find the answer in them, say: not available in the sources."
  4. 4Test it with twenty questions whose answers you know. A step with no shortcut — by it alone do you learn whether to trust it.

Where it errs

Three failures come up again and again. The first is a retrieval error, where it fetches the wrong passage and then answers from it confidently. The second is fragmented context, where it takes a clause from one page and leaves the exception to that clause sitting on the next. The third is an outdated version, which happens because nobody deleted the document that was superseded. This is why a knowledge space has to be reviewed periodically, exactly as any archive does.

Retrieve, then generate

Hallucination is the natural result of a statistical engine. The cure is to stop it answering from memory alone.

FIG. B17 — The path of a grounded answer
The question
⟶
Retrieval
⟶
The model
⟶
The answer + its grounding
Without grounding
a confident answer with no source, and an error found only after the damage.
With grounding
"I cannot find this in your sources" is an acceptable answer — and worth more than an invented one.

Without grounding: a confident answer with no source, and an error discovered only after the damage. With grounding: "I cannot find this in your sources" is an acceptable answer — and worth more than an invented one.

Do this

  1. 1 — On the tool. Build an assistant on ten documents from your work — not confidential ones — and write it the governing instruction. This is the longest exercise in the programme and the most useful.

  2. 2 — On the tool. Test it with twenty questions whose answers you know. Record the hit rate and the kind of errors: wrong retrieval, fragmented context, or an old version?

  3. 3 — In your field. Write the list of documents you will never upload, and why. That list matters more than the list of what you will.

What becomes possible after this unit

Once your archive can be questioned, three things change in your working day. The time you used to spend hunting for a document you knew existed disappears. Training a new employee gets faster, because they ask the archive before they ask you. And consistency becomes possible, because the answer is drawn from one source rather than from the memories of several people.

But keep one limit firmly in mind: a knowledge assistant reminds you; it does not decide for you. It brings back what you wrote yourself, and it knows nothing you never wrote down — and most of what actually governs your professional decisions is written in no file anywhere.

When to build and when to buy

Starting from scratch is a last option, not a first. One question settles it: is your problem common or singular?

FIG. B18 — The decision tree
Is your problem common or singular?
▼
Common
A pre-trained model from a model garden + fine-tuning on your data.
Hours instead of months · examples: image classification, sentiment analysis, translation
Singular and well-structured
Automated machine learning: you supply the data, the system tries the algorithms and chooses.
No data-science team needed · clean, organised data is
Singular and at scale
Building from scratch — the costliest option in time, compute and data.
Do not enter it unless the model itself is your product
Three traps: the cost of overkill (a vast model for a simple calculation) · the struggle with legacy systems that have no interfaces · the cloud-bill trap at scale.

Three traps: the cost of overkill — a vast model for a simple calculation · legacy systems with no interfaces to speak to · the cloud-bill trap at scale. And work out the unit economics from day one: if the value of the service does not cover the cost of running the model per user, success itself is what will sink you.

Where to after this unit? You have built the tool. The next unit brings it down into your own field — with the red lines that are never crossed.