backend/utils/web.py

"""
web.py — Real-time web search utilities.

Search cascade:
  1. Tavily  (fast, structured results, needs API key)
  2. DuckDuckGo  (free, no key needed) — fallback when Tavily fails/empty
"""
from __future__ import annotations

import asyncio
import logging
import re
from typing import Any

logger = logging.getLogger(__name__)


# ── Detection heuristics ──────────────────────────────────────────────────────

# Keywords that strongly indicate the user wants live/current information.
_REALTIME_WORDS = re.compile(
    r"""(?ix)
    \b(
      latest | newest | recent | currently | right now | as of today |
      today | tonight | this morning | this evening | this week | this month | this year |
      yesterday | last night | last week | last month |
      breaking | live | ongoing | just (happened|announced|released|launched|dropped) |
      in \s+ the \s+ news | trending | viral |
      current | now | update | updates |
      score | result | results | winner | winners | standings |
      stock | crypto | bitcoin | ethereum | price | prices | rate | rates | exchange \s+ rate |
      weather | forecast | temperature | rain | storm | hurricane | earthquake |
      election | vote | votes | poll | polls |
      match | game | tournament | championship | cup | league | ipl | nba | nfl | premier \s+ league |
      war | conflict | attack | protest | summit | deal | treaty |
      launch | released | announced | unveiled | dropped | update |
      who \s+ won | who \s+ is \s+ winning | what \s+ happened | what \s+ is \s+ happening |
      how \s+ much \s+ is | what \s+ is \s+ the \s+ (current|latest|price|value) |
      did \s+ \w+ \s+ (win|lose|beat|happen|say|announce|release) |
      when \s+ did | when \s+ is \s+ (the \s+)?(next|upcoming) |
      2024 | 2025 | 2026
    )\b
    """,
)

# Phrases that only make sense for live data (even without trigger words)
_REALTIME_PHRASES = [
    "current president",
    "prime minister",
    "today's",
    "this week's",
    "live score",
    "box office",
    "number one",
    "top charts",
    "inflation rate",
    "gdp growth",
    "unemployment rate",
    "oil price",
    "gold price",
    "market cap",
    "nifty 50",
    "sensex",
    "dow jones",
    "s&p 500",
    "nasdaq",
]

# Questions that are clearly NOT real-time (don't over-search)
_SKIP_RE = re.compile(
    r"""(?ix)
    \b(
      what \s+ is \s+ (machine \s+ learning|artificial \s+ intelligence|python|java|
                       gravity|photosynthesis|democracy|capitalism|blockchain|
                       deep \s+ learning|neural \s+ network) |
      how \s+ does \s+ \w+ \s+ work |
      explain \s+ (me \s+)? (the \s+)? \w+ |
      write \s+ (a|an|me) \s+ (poem|story|essay|function|code|program|script|letter) |
      help \s+ me \s+ (with|understand|write|debug|fix|code) |
      what \s+ are \s+ the \s+ (best \s+ practices|benefits|advantages|disadvantages) |
      history \s+ of |
      definition \s+ of |
      difference \s+ between
    )\b
    """,
)


def should_web_search(user_message: str) -> bool:
    """
    Decide whether the user message warrants a real-time web search.
    Returns True when the query is likely time-sensitive or current-events related.
    """
    msg = (user_message or "").strip()
    if not msg:
        return False

    # Skip clearly non-real-time educational/coding questions
    if _SKIP_RE.search(msg):
        return False

    # Hard-match real-time keywords
    if _REALTIME_WORDS.search(msg):
        return True

    # Phrase match
    msg_lower = msg.lower()
    if any(phrase in msg_lower for phrase in _REALTIME_PHRASES):
        return True

    # Question starting with "who is" / "who was" often needs current info
    if re.match(r"(?i)^who\s+(is|are|was|were)\s+", msg):
        return True

    # "What is X?" for things that change (stock prices, scores, etc.)
    if re.match(r"(?i)^what('?s|\s+is|\s+are)\s+(the\s+)?(current|latest|today)", msg):
        return True

    return False


# ── Tavily search ─────────────────────────────────────────────────────────────

async def _tavily_search(api_key: str, query: str, max_results: int) -> list[dict[str, Any]]:
    def _do() -> dict[str, Any]:
        from tavily import TavilyClient
        return TavilyClient(api_key=api_key).search(query=query, max_results=max_results)

    result = await asyncio.to_thread(_do)
    return result.get("results") or []


# ── DuckDuckGo search ─────────────────────────────────────────────────────────

async def _ddg_search(query: str, max_results: int) -> list[dict[str, Any]]:
    def _do() -> list[dict[str, Any]]:
        from duckduckgo_search import DDGS
        with DDGS() as ddgs:
            return list(ddgs.text(query, max_results=max_results))

    return await asyncio.to_thread(_do)


# ── Format results to context string ─────────────────────────────────────────

def _format_results(results: list[dict[str, Any]], source: str) -> str:
    if not results:
        return ""
    lines: list[str] = [f"🔍 Web search results ({source}):"]
    for i, r in enumerate(results, start=1):
        title   = (r.get("title") or "").strip()
        url     = (r.get("url") or r.get("href") or "").strip()
        snippet = (r.get("content") or r.get("body") or r.get("snippet") or "").strip()
        if not snippet:
            snippet = "No snippet available."
        lines.append(f"\n{i}. **{title}**\n   {snippet}\n   Source: {url}")
    return "\n".join(lines)


# ── Public entry point ────────────────────────────────────────────────────────

async def fetch_web_context(
    query: str,
    *,
    tavily_api_key: str | None = None,
    max_results: int = 5,
) -> str | None:
    """
    Try Tavily first, fall back to DuckDuckGo.
    Returns a formatted context string or None if both fail.
    """
    # 1. Tavily
    if tavily_api_key:
        try:
            results = await _tavily_search(tavily_api_key, query, max_results)
            if results:
                logger.info("web_search source=tavily query=%r results=%d", query, len(results))
                return _format_results(results, "Tavily")
            logger.info("web_search tavily returned 0 results for %r, trying DuckDuckGo", query)
        except Exception as exc:
            logger.warning("web_search tavily_failed=%r, falling back to DuckDuckGo", exc)

    # 2. DuckDuckGo fallback
    try:
        results = await _ddg_search(query, max_results)
        if results:
            logger.info("web_search source=duckduckgo query=%r results=%d", query, len(results))
            return _format_results(results, "DuckDuckGo")
    except Exception as exc:
        logger.warning("web_search duckduckgo_failed=%r", exc)

    return None


# ── Legacy compat (keep old name so graph.py import doesn't break) ────────────
async def tavily_search_to_context(
    api_key: str,
    query: str,
    *,
    max_results: int = 5,
) -> str:
    """Deprecated: use fetch_web_context instead."""
    ctx = await fetch_web_context(query, tavily_api_key=api_key, max_results=max_results)
    return ctx or "No relevant web results found."