3 RAG Projects to Build for Your AI Engineering Portfolio
If you want RAG projects that demonstrate more than a basic “chat with PDF” demo, build these three systems. Together they cover retrieval quality, multimodal document understanding, dynamic retrieval, verification, evaluation, and production-style failure handling.
Quick Overview
The three projects
Hybrid Search RAG
BM25 + dense vector retrieval + reranking + citation verification.
Multimodal Document RAG
PDFs + tables + scanned documents + OCR + structured extraction.
Agentic RAG
Dynamic retrieval + query reformulation + retry + self-correction.
Project 01
Hybrid Search RAG with Citation Verification
Build a RAG application that combines keyword retrieval with semantic vector search, reranks the candidate passages, generates an answer, and then verifies whether the cited sources actually support the claims in that answer.
Core architecture
Documents → parsing → chunking → metadata → dense embeddings + BM25 index → hybrid retrieval → candidate fusion → cross-encoder reranking → LLM generation → claim/citation verification → final answer.
Why it is valuable
Dense retrieval is strong for semantic meaning while BM25 can recover exact names, identifiers, error messages and terminology. Reranking focuses the final context, while citation verification adds a separate evidence check after generation.
Suggested implementation
1. Index: Split documents into meaningful chunks and preserve document, page and section metadata.
2. Retrieve: Run BM25 and vector search independently, then merge the candidate sets.
3. Rerank: Use a cross-encoder or another relevance model to reorder the top candidates.
4. Generate: Ask the LLM to answer only from the retrieved evidence and attach source references.
5. Verify: Break the answer into claims and check each claim against its cited passage.
6. Fail safely: If evidence is weak or missing, clearly flag the claim instead of presenting unsupported information as fact.
Metrics
Retrieval recall, precision@k, reranker relevance, citation support rate, answer faithfulness, latency and token cost.
Portfolio upgrade
Create a dashboard showing retrieval results, reranking scores, cited chunks, unsupported claims and before/after evaluation results.
Resume bullet
Built a hybrid-search RAG system with reranking and citation verification to improve retrieval accuracy and reduce unsupported answers.
Project 02
Multimodal Document RAG
Build a document intelligence pipeline that can work with normal PDFs, scanned pages, tables, figures and images instead of assuming every document is clean machine-readable text.
Pipeline
Upload → file classification → OCR where required → layout-aware parsing → table/image extraction → normalization → validation → chunking → indexing → retrieval → multimodal generation → confidence check → human review when needed.
What to handle
Invoices, reports, financial statements, manuals, forms, scanned contracts, tables, screenshots and pages containing mixed text and visual information.
Confidence-based validation
Do not treat extraction as correct simply because a model returned a value. Store field-level confidence, extraction method, source page and validation status. Route low-confidence or inconsistent results to a human review queue. For tables, validate row/column structure and totals when domain rules are available.
Skills demonstrated
OCR, document parsing, layout understanding, multimodal LLMs, structured extraction, validation and human-in-the-loop workflows.
Portfolio upgrade
Add a review interface where users can inspect the source page beside extracted fields and approve or correct low-confidence values.
Resume bullet
Built a multimodal document pipeline that extracts structured information from PDFs and tables with confidence-based validation and human review.
Project 03
Agentic RAG
Instead of always following the same retrieve-then-generate path, build an agentic system that can decide what information it needs, reformulate weak queries, retrieve again, inspect evidence quality, and only then generate the answer.
Example control loop
Question → plan → retrieve → inspect relevance → reformulate if weak → retrieve again → verify evidence → answer or ask for clarification. Keep explicit state so each decision is observable.
When it helps
Useful for ambiguous questions, multi-step research, large knowledge bases, terminology mismatches and tasks where a single retrieval pass frequently misses the needed evidence.
Design the agent as a state machine
Plan
Interpret the question and decide what evidence is required.
Retrieve
Search the knowledge base using the current query or sub-query.
Evaluate
Check whether the evidence is relevant and sufficient.
Correct
Rewrite, retry or request clarification before final generation.
Metrics
Task success, retrieval success, retry rate, answer faithfulness, latency, tool-call count and cost per task.
Portfolio upgrade
Show the agent trace: original query, reformulated query, retrieved evidence, decision, retry and final answer.
Resume bullet
Designed an agentic RAG system with dynamic retrieval, query reformulation, and self-correction for low-confidence results.
Suggested Technology Stack
Choose tools based on the problem
Retrieval
BM25, vector search, metadata filters, hybrid retrieval and reranking.
Vector databases
Qdrant, FAISS, Elasticsearch/OpenSearch or another production-appropriate store.
Document AI
OCR, layout-aware parsers, table extraction and multimodal model APIs.
Agent orchestration
A stateful workflow framework such as LangGraph, or your own explicit state machine.
API layer
FastAPI or another backend framework for ingestion, retrieval and generation endpoints.
Evaluation
Build a fixed test set and track retrieval relevance, faithfulness, correctness, latency and cost.
Build Plan
How to turn these ideas into real projects
Project Resources
Complete project details & preparation material
Use the project material below for implementation references, notes, examples and supporting resources.
Open Complete Project Details ↗If the shared folder requires access, sign in with the Google account that has permission. Availability of individual files may depend on the folder owner’s sharing settings.
Resume & Interview
How to explain these projects in an interview
Explain the problem first
Start with what failed in a basic RAG pipeline: missed exact terms, poor evidence, difficult documents or weak retrieval for ambiguous questions.
Explain the architecture
Walk through ingestion, indexing, retrieval, reranking, generation, verification and evaluation. Be able to explain the data flow without opening your code.
Explain failure handling
Give a concrete example where retrieval was weak and explain how your system detected the problem and recovered or refused to answer.
Explain the metrics
Know what you measured and why. Separate retrieval quality from generation quality instead of saying only that the chatbot “worked well.”
RAG interview checklist
Before putting these projects on your resume, make sure you can answer:
✓ What is RAG and why use it instead of fine-tuning?
✓ What is BM25 and when does keyword search help?
✓ What are embeddings and semantic similarity?
✓ How does hybrid retrieval combine different signals?
✓ Why use a reranker after initial retrieval?
✓ How do you select top-k?
✓ How can citations be verified rather than simply generated?
✓ How do OCR and layout affect document retrieval?
✓ How should low-confidence extraction be handled?
✓ What makes RAG agentic?
✓ How does query reformulation improve retrieval?
✓ How do you evaluate retrieval separately from answer generation?
✓ How do you reduce hallucinations and unsupported claims?
✓ How do latency and token cost change with retries and reranking?
Short version for your Instagram / Reel
3 RAG Projects to Build:
🔹 Hybrid Search RAG — BM25 + Vector Search + Reranking + Citation Verification
🔹 Multimodal RAG — PDFs + Tables + OCR + LLMs + Validation
🔹 Agentic RAG — Dynamic Retrieval + Query Reformulation + Self-Correction