src/config.py

"""Central configuration for the Medical RAG Chatbot.

Every other module imports its settings from here, so we only change
things in one place. Values can be overridden with environment variables.
"""
from pathlib import Path
import os

from dotenv import load_dotenv

# Load variables from a local .env file (e.g. GROQ_API_KEY) if present.
load_dotenv()

# --- Paths ---------------------------------------------------------------
PROJECT_ROOT = Path(__file__).resolve().parent.parent
DATA_DIR = PROJECT_ROOT / "data"          # source PDFs live here
CHROMA_DIR = PROJECT_ROOT / "chroma_db"   # persisted vector database

# --- Chunking (how documents are split before embedding) ----------------
# ~512 tokens with 50 overlap, per the briefing. We measure in characters
# here (~4 chars per token) which is what LangChain's splitter uses.
CHUNK_SIZE = 2000      # characters (~500 tokens)
CHUNK_OVERLAP = 200    # characters (~50 tokens)

# --- Embeddings (turns text into vectors) -------------------------------
# all-MiniLM-L6-v2 is small, fast, free, and runs on CPU.
EMBED_MODEL = os.getenv("EMBED_MODEL", "sentence-transformers/all-MiniLM-L6-v2")

# --- Retrieval -----------------------------------------------------------
TOP_K = 4              # how many chunks to retrieve per question

# --- LLM (Groq, free tier — Llama 3.3 70B) ------------------------------
LLM_MODEL = os.getenv("LLM_MODEL", "llama-3.3-70b-versatile")
LLM_TEMPERATURE = 0.1  # low = factual, less creative 
GROQ_API_KEY = os.getenv("GROQ_API_KEY")

# Name of the collection inside ChromaDB
COLLECTION_NAME = "medical_docs"