3 Resume-Ready RAG Projects: Hybrid Search, Multi-Modal Docs and Agentic RAG

Bashiri Smith · Facebook reel · 2026-09-15 · 0:05 · 40,911 views · Open on Facebook

Topics: Retrieval-Augmented Generation (RAG), Portfolio Projects, Resume, Job Search & Interviews · Level: intermediate

Summary

The video suggests three RAG portfolio projects you can build in a few hours, each with a resume bullet you can adapt. The projects are: (1) hybrid-search RAG with cross-encoder reranking and citation checks, (2) a multi-modal document pipeline for messy PDFs, tables and scanned files, with a human review step, and (3) agentic RAG with a retrieval router that can correct itself. Each one targets a quality or real-world gap that most simple RAG demos leave out.

Key points

• Project 1, hybrid-search RAG: combine dense vector search with BM25 keyword search, then rerank results with a cross-encoder.
• Add a citation-verification pass that checks whether each citation actually supports the claim it is attached to. The creator calls this the quality layer most RAG demos skip.
• Project 2, multi-modal document RAG: use OCR plus LLM extraction on messy PDFs, tables and scanned documents. The creator says this is what real enterprise data looks like, unlike clean markdown.
• Add a validation layer and a confidence-gated human review queue for low-confidence extractions.
• Project 3, agentic RAG: a retrieval router decides what to fetch and when, rewrites weak queries, and retries when the first pass returns too little instead of hallucinating an answer.
• The creator calls agentic RAG 'where the field is heading'.
• Frame each project as a resume bullet with a measurable result, for example 'cut retrieval errors and eliminated unsourced answers' or 'self-corrected on low-confidence results before generating an answer'.

Resources mentioned

• How to become an expert in RAG (BASWE AI Engineer Field Guide) · pdf · drive.google.com · free
  A one-page field guide from BASWE that lays out a 5-stage path to mastering retrieval-augmented generation.
  Also in: How RAG Works: Chunking, Embedding, Vector Storage, and Retrieval (Bashiri Smith on Facebook · notes), How RAG Works Under the Hood: Chunking, Embedding, Storage, Retrieval (Bashiri Smith on Facebook · notes)

Try this

• Build at least one of the three RAG projects and add the matching resume bullet to your resume.
• Add a citation-verification step to your RAG pipeline so it gives no unsourced answers.
• Test your pipeline on messy real-world documents such as scanned PDFs and tables, not clean markdown.
• Comment 'RAG' on the post to get the creator's build guides.
• Hybrid-search RAG combining dense vectors and BM25, with cross-encoder reranking and a pass that checks each citation supports its claim.
• Multi-modal document RAG: OCR plus LLM extraction from PDFs, tables and scanned documents, with a validation layer and a confidence-gated human review queue.
• Agentic RAG with a retrieval router that decides what to fetch, rewrites weak queries and retries when results are thin before answering.