ONE PROJECT THROUGHOUT
Build a document question-answering assistant
Give it a small set of documents you own or have permission to use. Ask a question. It finds the relevant passages, answers in plain language and shows which document supports the answer. If the documents do not contain the answer, it says so.
Keep version one small. Start with a few short, non-sensitive documents and one question at a time. A command-line prototype is enough. Add a web interface only when the answer flow works.
THE SIX SKILLS
Learn the next piece when you need it
Use this as an order of work, not a checklist to finish before building. Each milestone should leave you with something you can run or inspect.
THE FOUNDATION
Python
Learn enough to read a short program and change it confidently: variables, functions, lists and dictionaries, files, errors and packages. You do not need advanced algorithms to build the first version.
Build this
Write a script that reads a local text file, accepts a question and prints a response. Start with a simple keyword match so you understand the input and output before adding a model.
You are ready when
You can run the script, explain what each function does and fix a bad file path without starting over.
CONNECT & TRACK
APIs & Git
An API lets your program send a request to a model service and receive a response. Git records each useful change, so you can compare versions and recover a working one. Learn request and response shapes, authentication, environment variables, commits and branches.
Build this
Make one model API call from your Python script. Put the project in Git and commit the working version. Keep the API key in an environment variable, outside the repository.
You are ready when
You can show the request, read the response, explain a failed call and restore the last working commit.
MAKE THE MODEL USEFUL
LLMs & prompting
Learn what tokens and context windows limit, how instructions and examples shape an answer, and when structured output or a tool call is useful. A good prompt defines the task, supplies the right context and describes what to do when information is missing.
Build this
Send a question and one document excerpt to the model. Ask for a short answer, the supporting excerpt and a clear “not found” response when evidence is missing. Test questions with and without an answer.
You are ready when
The answer stays grounded in the supplied excerpt, and you can identify a case where it failed.
USE YOUR OWN KNOWLEDGE
RAG
Retrieval-augmented generation finds useful passages in your documents before asking the model to answer. Start with simple search over a few files. As the collection grows, learn chunking, embeddings and vector search. Always keep the source attached to each passage.
Build this
Split your documents into named passages, retrieve the most relevant ones and pass only those passages to the model. Show source names beside the answer.
You are ready when
You can test a known-answer question, a question across two documents and a question the documents cannot answer.
CONNECT THE STEPS
AI agents & automation
An agent can choose a tool or next step as it works toward a goal. First make the fixed question → retrieve → answer flow reliable. Then let the model call one bounded tool, such as document search, and inspect the result before it continues. Save only the state the workflow needs.
Build this
Add a search tool the model can call when it needs evidence. Log the tool request, retrieved passages and final answer. Keep a human review step for any action beyond answering.
You are ready when
You can trace a complete run and explain why the tool was called, what it returned and how the final answer used it.
MAKE IT USABLE
Database & deployment
A database stores documents, metadata and useful history after the program stops. Deployment makes the system available beyond your laptop. Learn basic tables, create/read/update operations, secrets, logs and how to keep the service running.
Build this
Store document titles and passages in a database. Put a small interface in front of the assistant and deploy it. Run the same three RAG test questions against the deployed version.
You are ready when
You can open a link from another device, ask a question, inspect its source and see a useful error if the service fails.
YOUR BUILD ORDER
From beginner to a first AI system
- 01Make it run locally.
Read one file, ask one question and print one response.
- 02Connect the model.
Call an API, keep your key private and commit a working version.
- 03Ground the answer.
Retrieve relevant passages, cite them and test missing answers.
- 04Add one tool.
Let the model search documents, then inspect the tool trace.
- 05Ship and test it.
Store the documents, deploy the app and try it from another device.
The goal is a small system you understand, can test and can improve—not a pile of tutorials completed.
START TODAY
Your first 30 minutes
Choose two short documents. Write three questions: one answered by document A, one by document B and one neither can answer. Then create a Python file that reads the documents and prints their titles. You have started building.
Help me build a beginner-friendly document question-answering assistant in Python. Teach me while we build it, one small step at a time. My starting point: [WHAT I KNOW ABOUT PYTHON] My documents: [TWO SHORT, NON-SENSITIVE DOCUMENTS] My three test questions: [ANSWER IN A], [ANSWER IN B], [NOT ANSWERED] Start with a local script that reads the files and prints their titles. Tell me exactly how to run it and how to tell if it worked. Then wait for my result before moving to the next step. Explain new code in plain language. When something fails, help me diagnose the exact error before adding features. Keep API keys out of code and Git. Later, help me add model calls, retrieval with source names, one search tool, a database and deployment. At every stage, give me one test for a known answer and one for a missing answer.