aiPublic proofVIRAL · 372+ REACTIONS
Academic RAG Assistant
Cited answers from lecture PDFs
Answers questions from academic documents with cited sources. Built for students to interact with lecture PDFs and research papers.
Case study snapshot
Role
RAG engineer
Audience
Students, researchers, and technical recruiters evaluating document AI skill
Timeline
2025-07
Role
RAG engineer
Audience
Students, researchers, and technical recruiters evaluating document AI skill
Date
2025-07
Problem
Why it mattered
Students spend too much time searching lecture PDFs and research notes, and generic chatbots often answer without citations.
Solution
What shipped
A RAG assistant that chunks academic PDFs, retrieves relevant context, and produces grounded answers with source references.
Impact
Proof signal
372+ LinkedIn reactions
< architecture />
System flow
PDFs are parsed into chunks, embedded, stored in a vector database, retrieved by semantic similarity, and passed to an LLM with citation instructions.
PDF Upload
Chunking
Embeddings
Vector Store
Query → LLM
Cited Answer
Visual architecture
01
PDF Upload
02
Chunking
03
Embeddings
04
Vector Store
05
Query → LLM
06
Cited Answer
quality score
Honest maturity snapshot
Code Quality8/10
UI / UX6/10
Scalability7/10
Production Ready6/10
Before
Manual questions, scattered context, and slow follow-up.
After
An agent workflow that routes intent, calls tools, and returns useful next actions.
capabilities
Features
- PDF upload and chunking pipeline.
- Embedding-backed semantic retrieval.
- Cited answers from lecture and research material.
- Streamlit interface for fast student testing.
- Project story optimized for public sharing.
tradeoffs
Engineering challenges
- Preventing hallucinated citations.
- Choosing chunk sizes that preserve enough academic context.
- Explaining RAG value clearly to non-specialist students.
proof
Results
- Generated 372+ LinkedIn reactions.
- Established strong public proof of RAG system understanding.
- Became a foundation for later agentic RAG projects.
reflection
What I learned
- Tradeoff learned: Preventing hallucinated citations.
- Proof learned: Generated 372+ LinkedIn reactions.
- Next iteration: Add per-answer citation confidence.
< stack />
Gemini APILangChainStreamlitPinecone
< next iteration />
Future improvements
Add per-answer citation confidence.
Support multi-document collections by course.
Add quiz generation and spaced-repetition study cards.