| # Textbook RAG Assistant (LangChain + Chroma + OpenAI) |
|
|
| End-to-end **Retrieval-Augmented Generation (RAG)** system built for **Applied AI Engineering** workflows: ingest PDF textbooks, chunk + enrich metadata, embed into a vector database, evaluate retrieval quality, and ship a lightweight **Gradio** UI for interactive Q&A. |
|
|
| **Live repo:** https://github.com/zainabahmed4626-lab/AIEngineeringWeek2V1 |
|
|
| --- |
|
|
| ## What this project demonstrates |
|
|
| - **Production-shaped RAG**: not “call an LLM”—a full pipeline from documents → chunks → embeddings → retrieval → grounded answers. |
| - **Grounding + safety posture**: answers are constrained to retrieved context with an explicit fallback when evidence is insufficient. |
| - **Retrieval engineering**: hybrid retrieval (**vector similarity + BM25**) and **metadata-aware filtering** hooks (source / section / date). |
| - **Evaluation discipline**: a small eval loop reporting retrieval / faithfulness / correctness style signals (assignment-oriented, extensible). |
| - **Product thinking**: a simple **Gradio** interface for manual testing, plus debug panels to inspect retrieved chunks. |
|
|
| --- |
|
|
| ## Architecture (high level) |
| <img width="1920" height="991" alt="image" src="https://github.com/user-attachments/assets/028fc97a-ca7f-4777-b332-59e8587d504f" /> |
|
|
| PDFs[PDF textbooks] --> Load[LangChain PDF loader] |
| Load --> Chunk[RecursiveCharacterTextSplitter + metadata] |
| Chunk --> Embed[HuggingFace embeddings] |
| Embed --> VS[Chroma vector store] |
| VS --> RetV[Vector retriever] |
| Chunk --> RetB[BM25 retriever] |
| RetV --> Ens[Ensemble retriever] |
| RetB --> Ens |
| Ens --> QA[RetrievalQA (OpenAI chat)] |
| QA --> UI[Gradio UI] |
| ``` |
| |
| --- |
| |
| ## Tech stack |
| |
| - **Orchestration**: LangChain (`RetrievalQA`, prompts, retrievers) |
| - **Embeddings**: `sentence-transformers/all-MiniLM-L6-v2` (via LangChain community integrations) |
| - **Vector DB**: Chroma (persistent store) |
| - **Hybrid retrieval**: BM25 (`rank-bm25`) + dense vectors (`EnsembleRetriever`) |
| - **LLM**: OpenAI chat model (`gpt-3.5-turbo`, temperature configurable) |
| - **UI**: Gradio |
| - **Notebook**: `rag1.ipynb` (single runnable artifact for the assignment) |
| |
| --- |
| |
| ## Key features implemented |
| |
| - **Step 1 — Load**: PDF ingestion + sanity prints (document count + preview) |
| - **Step 2 — Chunk**: `RecursiveCharacterTextSplitter` with two chunking experiments + stats |
| - **Step 3 — Embed + Store**: embeddings persisted to Chroma (`./chroma_db`, collection `textbook_rag`) |
| - **Step 4 — Retrieval test**: `similarity_search` with annotated relevance notes |
| - **Step 5 — RAG chain**: `RetrievalQA` + custom prompt + retriever tuning |
| - **Step 6 — Evaluation**: mini eval set + structured reporting + aggregate scores |
| - **Hybrid retrieval**: vector + BM25 ensemble for stronger keyword coverage |
| - **Metadata enrichment**: `source`, `section`, `date` on chunks + optional filtered retrieval path |
| - **UX polish**: greeting handling in the Gradio path for a friendlier chat experience |
| |
| --- |
| |
| ## Quickstart (local) |
| |
| ### 1) Prerequisites |
| |
| - Python 3.10+ recommended (your notebook metadata may show newer; adjust if needed) |
| - OpenAI API access |
| |
| ### 2) Configure secrets locally |
| |
| Create a `.env` file in the project folder: |
| |
| ```bash |
| OPENAI_API_KEY=...your key... |
| ``` |
| |
| This repo intentionally **does not** commit `.env`. |
| |
| ### 3) Add your PDFs |
| |
| Place these next to `rag1.ipynb` (not committed here): |
| |
| - `textbook_1.pdf` |
| - `textbook_2.pdf` |
| |
| ### 4) Run the notebook |
| |
| Open `rag1.ipynb` and run cells **in order** (setup → load → chunk → embed → retrieval tests → RAG → eval → optional UI). |
| |
| --- |
| |
| ## Observations — What Worked, What Didn’t, What I’d Improve |
| **What Worked** |
| The combination of metadata filtering + hybrid retrieval (BM25 + vector search) noticeably improved relevance, especially for definition‑heavy sections. |
| Chunking at 500 characters produced more accurate retrieval for specific questions. |
| The custom grounding prompt kept answers concise and reduced hallucinations. |
| |
| **What Didn’t Work** |
| Some textbook sections still produced noisy chunks due to headers, footers, and formatting artifacts. |
| A few queries returned context that was technically related but not the most precise match. |
| |
| **What I’d Improve** |
| Clean the PDF text further before chunking to remove repeated structural elements. |
| Experiment with stronger embedding models to improve semantic matching. |
| Add lightweight evaluation automation to speed up iteration. |
| |
| --- |
| |
| ## Optional helper script |
| |
| `build_github_push_payload.py` generates a JSON payload for GitHub uploads. It is **not** required to run the RAG notebook. |
| |
