Project guide
RAG portfolio projects that show real skill (with Indian data ideas)
RAG project ideas for an AI engineering portfolio, from beginner to advanced, with Indian datasets, what to measure and how to present them to employers.
Short answer: a strong RAG portfolio project answers questions over real, messy documents, cites its sources, and comes with numbers: retrieval recall@k, answer faithfulness and cost per query, before and after your improvements. A chatbot over one clean PDF shows you followed a tutorial; an assistant over many real documents with an evaluation report shows you can engineer.
What makes a RAG project stand out
- Real documents. Scanned pages, tables, several formats, long files.
- Citations. Every answer links to the passage it used.
- Evaluation. A small golden set of questions with recall@k, MRR and a faithfulness score.
- One improvement, measured. For example "hybrid search raised recall@5 from X to Y on my test set", with your own numbers.
- A clean README. Problem, data, architecture diagram, results table, how to run it, limitations.
Project ideas, from first to advanced
| Level | Project | Data idea | What to measure |
|---|---|---|---|
| Beginner | Q&A over a public policy document with citations | A public Indian government scheme guideline PDF | Does each answer cite the right section? |
| Intermediate | Bilingual assistant (English and Hindi) | Public government documents available in both languages | Recall@k per language; answer faithfulness |
| Intermediate | Hybrid search over product or legal text with codes | Public notifications with section numbers and form names | Exact-code queries vs paraphrase queries |
| Advanced | Text-to-SQL plus RAG | An open dataset in a database plus its documentation | SQL accuracy on a test set |
| Advanced | RAG with an evaluation suite in CI | Any of the above | Pass rate per release; regressions caught |
Use only documents you are allowed to use (public government documents and open datasets are good choices), and do not upload personal data.
A template README section
Results (my test set of N questions). Keyword only: recall@5 = A. Vector only: B. Hybrid + reranker: C. Faithfulness (LLM judge, calibrated on 50 human labels): D. Median latency: E s. Cost per 1,000 queries: ₹F.
Replace the letters with your own measured numbers; never present made-up results.
Mistakes that weaken a project
- No evaluation at all, only screenshots of good answers.
- Secrets (API keys) committed to the repository.
- No mention of limitations or failure cases.
- A demo that no longer runs because the free API key expired: record a short demo video too.
The project inside the program
Module 8 (RAG Engineering) ends with a bilingual policy assistant: RAG over public Indian government documents in English and Hindi, with citations and a retrieval-quality report. Module 9 then adds an evaluation suite and red-team report wired into CI. These are curriculum projects; see all 12 projects. Related: RAG explained · Hybrid search guide · RAG engineering course.