All micro-internships
Micro-internship brief ≈17h across 2 daysbuilder

Build a mini app that chats with your documents

Retrieval-augmented generation (RAG) is used across modern AI products and roles. Build an app that answers questions from 3+ real documents with citations, and prove it refuses questions the documents can’t answer.

AI Application DevelopmentPrompt EngineeringResearch & Fact-checkingPython for DataRetrieval-Augmented Generation

Posted by The AI Internships

The work brief

  1. 01Pick 3+ real documents you care about (syllabus PDFs, course prospectuses, textbook chapters, or game rules).
  2. 02Build a doc-Q&A app: code it with AI help (LangChain/LlamaIndex or plain embeddings), or use Flowise/Langflow — but you must wire it yourself, not just upload to NotebookLM.
  3. 03Make answers cite which document/section they came from.
  4. 04Test with 6 questions: 4 answerable from the docs, 2 traps that aren’t. Push the code to a public repo.

What you’ll produce

4 deliverables

Submission standard

Submit your docs description, the public repo, a screenshot of it running, and the 6 test transcripts (including trap behaviour). Include your AI workflow: tools, best prompts, and what you changed from the AI’s output. Your repo's commit history must show the pipeline coming together over more than one sitting, not a single upload.

  • The documents you used and why

    Written responseRequired

  • Public GitHub repo with your code/flow export

    Public GitHub repo with your code/flow export — commit history should span multiple sessions (chunking, retrieval, prompt tweaks as separate dated commits), not one paste. A single dump commit fails the authenticity gate.

    Public linkRequired

  • Screenshot of the app answering with a citation

    Image uploadRequired

  • 6 test Q&As: 4 answerable (with citations), 2 traps — plus what you fixed

    RAG means retrieval-augmented generation. Include the source citation for answerable questions and show that trap questions are refused or clearly marked unsupported.

    Written responseRequired

You’ll complete these inside your private workspace.

How your work is evaluated

The passing benchmark is 70/100.

Working RAG

29%

The app answers from the documents with citations.

Trap handling

29%

Trap questions get honest "not in the documents" responses.

Built, not uploaded

14%

The repo shows they wired the pipeline themselves.

Wired over time, not pasted

29%

If the repo is a single copy-paste commit, is private/empty, or the running screenshot doesn't match the repo code, this fails regardless of transcript prose.

How we grade your AI usage

30% of your score

Using AI is the point — it’s the skill this certificate proves. You’ll answer three short questions about how you used it: what you asked, what was wrong with its first answer, and what you changed. Specific, honest answers score high. “I pasted the brief and submitted the answer” scores near zero.