config.py

import os

#  Embedding 
EMBEDDING_MODEL = "all-MiniLM-L6-v2"   # Fast, lightweight, great for semantic search

#  LLM 
LLM_MODEL = "google/flan-t5-base"       # Instruction-tuned, actually answers questions on CPU

# Chunking 
CHUNK_SIZE = 500
CHUNK_OVERLAP = 100
MIN_CHUNK_SIZE = 100                    # Skip tiny useless chunks

# Retrieval 
TOP_K = 3

#  Generation
MAX_NEW_TOKENS = 200
TEMPERATURE = 0.2                       # Low = more factual, less random

#  Storage 
VECTOR_STORE_PATH = "vectorstore/"

#  App 
DEVICE = "cpu"
SUPPORTED_FORMATS = [".txt", ".pdf"]
MAX_FILE_SIZE_MB = 10