How calebagoha123/jonathan_ta works
Main Components
a. Document Processing Pipeline
• Source PDFs (in src/web/static/docs/)
• document_loader.py (ingests raw PDFs)
• preprocessor.py (parses, text-chunks, attaches metadata)
b. Vector Storage
• ChromaDB instance (data/chroma_db/)
• Collections per document type
• chroma.sqlite3 for persistence
c. Embedding & Retrieval Layer
• Sentence-Transformer model (all-MiniLM-L6-v2)
• Embeddings module (src/rag/embeddings.py)
• Retriever (src/rag/retriever.py)
• Semantic search + metadata filtering
d. LLM & Prompting
• External LLM endpoint (gpt-4o-mini)
• Query formatting / prompt assembly
e. Web Interface
• Streamlit app (src/web/streamlit_app.py)
• Templates (src/web/templates/index.html)
• Static assets & course PDFs
• Session/context management
Architectural Patterns
• Layered architecture: Ingestion → Data store → Retrieval/Compute → Presentation
• RAG pattern: Retrieval augmented generation with vector DB + LLM
• Modularity: Separate concerns into src/data, src/rag, src/web
Technologies & Frameworks
• PDF parsing & preprocessing: PyPDF2 or similar (in document_loader.py)
• Chunking logic (1000 tokens/100 overlap)
• Vector DB: ChromaDB (local embedded)
• Embeddings: Sentence Transformers (all-MiniLM-L6-v2)
• LLM: OpenAI gpt-4o-mini via API
• Web: Streamlit, Jinja2 (templates), CSS/HTML static assets
• Dev environment: Docker devcontainer