config.py

import os
from pathlib import Path

ROOT = Path(__file__).resolve().parent.parent

DB_DIR = os.getenv("IPBOT_DB_DIR", str(ROOT / "vector_db"))
KB_DIR = os.getenv("IPBOT_KB_DIR", str(ROOT / "Knowledge_Base"))

EMBED_MODEL = "intfloat/multilingual-e5-small"
LLM_MODEL = os.getenv("IPBOT_LLM_MODEL", "llama-3.3-70b-versatile")

CHUNK_SIZE = 1200
CHUNK_OVERLAP = 120

# two searches per question: one on the rewritten query, one on what the student actually typed
# the pool is re-ranked and trimmed to MAX_CONTEXT_DOCS
# 10 chunks is roughly 12k characters, which llama 3.3 swallows without noticing and which is usually the difference between half an answer and a whole one
SEARCH_K_ROUTED = 14
SEARCH_K_RAW = 6
MAX_CONTEXT_DOCS = 10

ROUTER_HISTORY_TURNS = 6
ANSWER_HISTORY_TURNS = 16

MAX_MESSAGE_CHARS = 2000
MAX_TURN_CHARS = 4000
MAX_HISTORY_TURNS = 40
MAX_DOC_CHARS = 6000
MAX_UPLOAD_BYTES = 5_000_000

TIMEZONE = "Europe/Lisbon"