Home / Case Studies / AI Learning Assistant
Case study / Retrieval-Augmented Generation

Stop guessing.
Cite the page.

As a TA, I answered the same three CNN/backpropagation questions a hundred times. So I built an AI that answers them instead — one that points at the exact textbook page it got the answer from, instead of confidently making one up.

RoleSolo build
StackLangChain, ChromaDB, Groq
ModelLlama-3.3-70B-versatile
LinksGitHub repo →
AI Learning Assistant pipeline: question to cited answer Student question → ChromaDB retrieval MMR + filtering BGE embeddings → Guardrail on-topic filter → Llama-3.3-70B via Groq → Cited answer source + page #
AI Learning Assistant pipeline — question to source-cited answer

The itch

Running PyTorch labs and BERT fine-tuning sessions as a Teaching Assistant at IIIT Lucknow put me in front of the same conceptual confusion, over and over, from different students: why does backprop break on this architecture, what's actually different between an RNN and an LSTM cell, why is this fine-tune overfitting after two epochs. The questions were legitimate. Answering the same one for the fifteenth time wasn't a good use of anyone's time — mine or theirs, since a live queue means everyone waits.

The obvious fix, "just point them at ChatGPT," has an obvious problem: general-purpose LLMs answer confidently whether or not they're right, and a student who doesn't yet know the material has no way to tell a correct explanation from a plausible-sounding wrong one.

What already existed

Generic LLM chatbots are fluent but ungrounded — ask one a subtle ML question and it will produce a clean, well-formatted, occasionally wrong answer with the same confident tone as a right one. There was no shortage of RAG demo projects online either, but most of them stopped at "retrieve some chunks, stuff them in a prompt" without the parts that actually make a RAG system trustworthy in an academic setting: source attribution down to the page, retrieval that favors diverse relevant context over near-duplicate chunks, and guardrails to stop it wandering off-topic.

A student who can't verify an answer against the actual source hasn't learned the material — they've just outsourced their doubt to a chatbot.

How I actually built it

I built the knowledge base from two authoritative ML/DL textbooks rather than the open web, so every answer is grounded in material the student can actually go check. Text gets embedded with BAAI/bge-base-en-v1.5 and stored in ChromaDB for semantic search.

For retrieval, I didn't settle for naive top-k similarity search. I layered MMR (Maximal Marginal Relevance) to keep retrieved chunks diverse instead of five near-duplicates of the same paragraph, plus an embedding-filtering step to drop chunks that pass a similarity threshold but aren't actually relevant to the question. Groq's Llama-3.3-70B-versatile generates the final answer — chosen for the inference speed, since a slow answer defeats the purpose in a live lab session — and every response comes back with the source and page number it was pulled from.

Guardrails, not just generation

I added query-filtering guardrails specifically so the assistant rejects questions outside its knowledge base instead of hallucinating an answer anyway, plus a WindowBufferMemory for conversation memory, so a student can ask a follow-up without re-explaining their original question.

What it does

  • Answers ML/DL concept questions with source citations and page numbers, not unverifiable prose
  • MMR retrieval + embedding filtering instead of naive top-k similarity search
  • Guardrails that reject off-topic queries rather than answering anyway
  • Conversation memory for natural multi-turn follow-ups
  • Built on Streamlit for a lightweight, fast-to-iterate interface

Why it matters

This wasn't built as a portfolio flex — it came directly out of a real bottleneck in my own TA work. Cutting the repeat-question load frees up TA time for the questions that actually need a human, and giving students a citation instead of an unverifiable answer is the difference between a study tool and a liability.

FAQ

What is the AI Learning Assistant?

A retrieval-augmented generation (RAG) tool that answers ML/DL concept questions grounded in two authoritative textbooks, citing the source and page number behind every answer instead of generating unverifiable prose.

How does it avoid hallucinating wrong ML answers?

It only answers from a fixed knowledge base of two ML/DL textbooks retrieved via ChromaDB, uses MMR retrieval plus embedding filtering to keep context relevant and diverse, and has query-filtering guardrails that reject questions outside its knowledge base rather than answering anyway.

What LLM powers the AI Learning Assistant?

Groq's Llama-3.3-70B-versatile, chosen for fast inference speed suited to live lab sessions, with BAAI/bge-base-en-v1.5 embeddings for retrieval.

Why did Aishrica Dhiman build this?

As a Teaching Assistant at IIIT Lucknow running PyTorch and BERT fine-tuning labs, she kept fielding the same conceptual questions from different students. The assistant was built to reduce that repeat-question load and give students source-cited answers they can verify themselves.

← Previous: SmartPlate Next case study: Saarthi-AI →