Artificial Intelligence
Build an Azure RAG project in 7 days
Build a portfolio-ready, citation-backed knowledge assistant with a seven-day plan for retrieval, grounding, evaluation, security, and documentation.
Documentation checked: verified against the official references below on October 13, 2026. Product names, preview status, and certification objectives can change.
Interactive project challenge
Your seven-day build plan
Check off each day as you finish its evidence. Your progress is saved in this browser so you can return to the project later.
Your project progress
0 of 7 days complete
0 of 7 challenge days complete.
A small, testable knowledge assistant for one clearly defined audience.
- Choose a low-risk scenario and a small set of public or synthetic documents.
- Write 10 realistic questions and identify the source passage each should use.
- Define when the assistant must say it does not know, plus a latency and cost target.
Evidence: A one-page project brief, source list, and question set.
Documents can be retrieved with enough context to support useful citations.
- Keep stable document IDs and record each source title, section, URL, and version.
- Choose chunk boundaries that preserve headings and meaning instead of splitting blindly.
- Write down how updated or deleted source documents will be reflected in the index.
Evidence: A sample of prepared chunks and a short indexing decision note.
You can inspect whether the expected evidence appears in search results.
- Create a basic keyword-search path over the prepared content.
- Run the 10 questions from Day 1 and record the returned source IDs and ranks.
- Calculate Recall@5: the share of questions where an expected source appears in the first five results.
Evidence: A baseline result table with misses and false matches called out.
A measured retrieval choice, not a search mode selected by habit.
- Add vector retrieval and compare it with the keyword baseline using the same questions.
- Try hybrid retrieval, which combines keyword and vector results, then inspect ranking changes.
- Record quality, latency, and any extra service or embedding requirements.
Evidence: A before-and-after table and a justified retrieval choice.
Answers are traceable to retrieved evidence and can abstain when evidence is missing.
- Pass only the most relevant passages into the model request.
- Return source links or identifiers that let a reader verify each important claim.
- Treat retrieved text as untrusted data, not as instructions, and test a question with no supported answer.
Evidence: Three example answers: supported, partially supported, and correctly unanswered.
A repeatable test report exposes failures before the project is presented as reliable.
- Score expected-source retrieval, citation support, answer usefulness, and abstention separately.
- Add ambiguous, outdated, conflicting, unauthorized, and prompt-injection test cases.
- Use Microsoft Entra ID and least privilege for Azure access; keep credentials out of source control.
Evidence: A small evaluation report with failures, fixes, and remaining limitations.
Another engineer can understand what you built, how you tested it, and what remains unfinished.
- Create a README with setup, architecture, test method, and a short demo walkthrough.
- Document one trade-off, one observed failure, and the next improvement you would make.
- Review current Azure pricing, set a budget before deployment, and remove test resources when finished.
Evidence: A shareable repository or case study with architecture, evaluation results, and cleanup steps.
Progress stays in this browser on this device. It is not sent to Migo Mind.
The project you will build
Build a small knowledge assistant that searches a bounded collection of public or synthetic documents, answers with inspectable source references, and says when its evidence is not enough. The goal is a clear engineering demonstration, not a production service or a chat interface for its own sake.
- A defined audience and a narrow knowledge task
- A repeatable set of questions with expected source passages
- A retrieval baseline compared with a justified alternative
- Grounded responses, citations, and an explicit no-answer behavior
- A concise evaluation report and portfolio-ready README
Before you create Azure resources
You can complete the project brief, source preparation, test design, and portfolio plan without deploying cloud resources. If you choose to use Azure AI Search or a model deployment, check current regional availability, quotas, pricing, and service requirements first. Set a budget alert, use non-sensitive sample content, and remove resources when the experiment is finished.
- Do not put credentials, customer documents, or personal information in a public repository
- Use Microsoft Entra ID and least-privilege roles where supported
- Keep retrieval permissions in the application and service design; do not rely on model instructions to enforce access
- Treat preview features as experiments and document them clearly
Turn the challenge into portfolio evidence
A reviewer should be able to see what problem the system solves, how you know its answers are supported, and what still fails. Publish a short walkthrough with an architecture diagram, a small redacted test report, and one honest limitation. Never claim production readiness from a learning project.
- Describe the user need and success criteria
- Show the ingestion, retrieval, model, identity, and citation flow
- Report retrieval and answer evaluation separately
- Explain one quality, security, latency, or cost trade-off
- State cleanup steps and the next improvement
Choose your next step
Use the project as a learning loop. Keep improving it when your tests show a clear reason.
Read the Azure RAG guide
Review the system boundaries, retrieval choices, security, and operational checks behind the challenge.
Open resource EvaluateStudy agent evaluation
Apply repeatable quality, safety, latency, and cost checks to the behavior you build.
Open resource CareerPrepare to explain the project
Practice communicating Azure AI design choices in a role-based interview format.
Open resourceFrequently asked questions
Do I need an Azure subscription to do the challenge?+
No. You can complete the planning, source analysis, question set, and evaluation design without one. Azure deployment is optional and may incur charges, so check current pricing and clean up resources.
How much time does each day take?+
Plan for a focused work session each day, but treat the seven days as an order of work rather than a deadline. Keep the test collection small and expand only when the evidence shows a need.
Is this a production-ready reference implementation?+
No. It is a learning challenge for producing a tested portfolio project. Production systems need workload-specific security, privacy, availability, monitoring, and operational reviews.
Where is my challenge progress saved?+
The checked days are stored in local browser storage on your current device. They are not sent to Migo Mind, and clearing site data removes them.
Official sources
Use these primary references to verify product capabilities, exam objectives, and implementation details.
Put this into practice
Use the production-minded RAG guide for architecture decisions and implementation context.
Open resource