Enterprise Knowledge Assistant — Advanced RAG Capstone

An enterprise multi-document RAG (Retrieval-Augmented Generation) system. Upload PDF / DOCX / TXT / MD / CSV files, ask a question, and get an answer grounded in your own documents — with hybrid search and reranking.

Architecture

Streamlit Dashboard
       |
  Upload Files ---------------- Ask Question
       |
LlamaIndex Document Readers
       |
PDF / DOCX / TXT / MD / CSV Parsing
       |
    Chunking
       |
OpenAI Embeddings (text-embedding-3-small)
       |
   Chroma Vector Store
       |
   Hybrid Retrieval (Vector Search + BM25)
       |
   BGE Reranker (Free, local)
       |
   Top Relevant Chunks
       |
   OpenAI GPT-4.1 mini
       |
   Final Answer

Files and which stage they handle

File	Architecture stage
document_loader.py	LlamaIndex Document Readers + file parsing
chunking.py	Chunking
embeddings_store.py	OpenAI Embeddings + Chroma Vector Store
hybrid_retriever.py	Hybrid Retrieval (Vector + BM25)
reranker.py	BGE Reranker
generator.py	OpenAI GPT-4.1 mini -> Final Answer
app.py	Streamlit Dashboard (connects every stage)
config.py	All settings in one place

Setup

1. Create a virtual environment (recommended):

   python -m venv venv
   venv\Scripts\activate      (Windows)
   source venv/bin/activate   (Mac/Linux)
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Run the app:

   streamlit run app.py
   

4. Enter your OpenAI API key in the sidebar and click Set. The key is stored only for the current Streamlit session.

How it works, in plain words

1. Upload — You upload one or more documents in the browser.
2. Read & Parse — LlamaIndex opens each file, no matter the format, and pulls out plain text.
3. Chunk — Long text is cut into small overlapping pieces so the AI can search and read them efficiently.
4. Embed & Store — Each chunk is turned into a list of numbers (an embedding) using OpenAI's cheapest embedding model, and saved in Chroma, a free local vector database.
5. Hybrid Retrieval — When you ask a question, we search two ways:
   • Vector search: finds chunks with similar meaning.
   • BM25: finds chunks with matching keywords. Both lists are merged using Reciprocal Rank Fusion (RRF).
6. Rerank — A free local model (BGE reranker) looks closely at the question and each candidate chunk together, and picks the best few.
7. Generate — The best chunks are sent to OpenAI's gpt-4.1-mini along with your question, and it writes the final answer.

Cost notes

This project only uses cheap OpenAI models:

• text-embedding-3-small for embeddings
• gpt-4.1-mini for the final answer

Chroma (vector database) and the BGE reranker both run for free on your own computer — no extra API cost for search or reranking.