src/config.py

# Central config — reads everything from .env so we can swap
# LLM providers, tweak retrieval params, etc. without touching code

import os
from pathlib import Path
from dotenv import load_dotenv

load_dotenv()

# -- paths --
ROOT_DIR = Path(__file__).resolve().parent.parent
DATA_DIR = ROOT_DIR / "data"
RAW_DATA_DIR = DATA_DIR / "raw"
PROCESSED_DATA_DIR = DATA_DIR / "processed"
CHROMA_DIR = PROCESSED_DATA_DIR / "chroma_store"
DUCKDB_PATH = PROCESSED_DATA_DIR / "financials.duckdb"
BM25_INDEX_PATH = PROCESSED_DATA_DIR / "bm25_index.pkl"

# -- LLM setup --
LLM_PROVIDER = os.getenv("LLM_PROVIDER", "openai").lower()
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", "")
ANTHROPIC_API_KEY = os.getenv("ANTHROPIC_API_KEY", "")
GROQ_API_KEY = os.getenv("GROQ_API_KEY", "")

# default models per provider — override with LLM_MODEL in .env
_DEFAULT_MODELS = {
    "openai": "gpt-4o-mini",
    "anthropic": "claude-sonnet-4-20250514",
    "groq": "llama-3.3-70b-versatile",
}
LLM_MODEL = os.getenv("LLM_MODEL") or _DEFAULT_MODELS.get(LLM_PROVIDER, "gpt-4o-mini")

# embedding model — runs locally, no API key needed
# use "or" instead of default param so empty string in .env still falls back
EMBEDDING_MODEL = os.getenv("EMBEDDING_MODEL") or "all-MiniLM-L6-v2"

# -- retrieval tuning --
SIMILARITY_THRESHOLD = float(os.getenv("SIMILARITY_THRESHOLD", "0.35"))
TOP_K_RESULTS = int(os.getenv("TOP_K_RESULTS", "10"))
RERANK_TOP_N = int(os.getenv("RERANK_TOP_N", "5"))
CHUNK_SIZE = 1500       # chars per chunk — bigger chunks = fewer total = faster indexing
CHUNK_OVERLAP = 100     # overlap between consecutive chunks

# -- logging --
LOG_LEVEL = os.getenv("LOG_LEVEL", "INFO")

# -- dataset info (shown to the planner agent + UI) --
DATASET_DESCRIPTION = """
Available structured data (DuckDB — SQL queries):
  Table: companies
    Columns: company_name, ticker, sector, revenue_mn (in millions USD),
             net_income_mn, total_assets_mn, market_cap_mn, year, quarter
    Coverage: ~5,000 publicly traded US companies, 2020–2024 quarterly data

Available unstructured data (ChromaDB + BM25 — semantic & keyword search):
  Source: Financial news articles and analyst reports
  Fields per chunk: text, source, date, companies_mentioned, sector
  Coverage: ~100,000 articles from 2020–2024
  Topics: earnings, market analysis, sector trends, M&A, analyst opinions
""".strip()