Blame for news/relevance.py

"""Business-relevance pre-filter (DataWiz-Project's `pipeline/relevance.py` pattern,
applied to MSME news instead of security incidents).

Groq is the expensive step of this pipeline, not scraping: every scraped item that
reaches classify_item() costs an API call, and a feed mixes in sports, entertainment,
school-life and aviation-incident filler alongside the genuinely business-relevant
items. This module is a cheap, zero-cost keyword check that runs first — an item
with no match to any business/economy/regulatory signal never reaches Groq at all
and is auto-rejected on the spot, the same way DataWiz discards a zero-match article
before it ever reaches the NLP extractor.

This is intentionally coarser than DataWiz's incident classifier: it only answers
"is this worth spending a Groq call on", not what category/impact it is — that fine
classification still happens in classify.py. A false positive here just costs one
wasted Groq call (same as today); a false negative would silently drop a relevant
item, so the keyword list stays broad on purpose.
"""

import re

# Roots covering the five categories classify.py sorts into, plus general
# economic/business-operational signal. Stored as roots, matched as `\b<root>\w*\b`
# so one entry covers inflected forms (e.g. "levy" matches "levies").
BUSINESS_KEYWORDS: dict[str, str] = {
    # policy / regulatory
    "tax": "policy", "vat": "policy", "levy": "policy", "tariff": "policy",
    "duty": "policy", "customs": "policy", "excise": "policy", "regulat": "policy",
    "cbn": "policy", "naira": "policy", "forex": "policy", "fx rate": "policy",
    "exchange rate": "policy", "interest rate": "policy", "monetary polic": "policy",
    "minimum wage": "policy", "subsidy": "policy", "fuel price": "policy",
    "petrol price": "policy", "firs": "policy", "son ": "policy", "nafdac": "policy",
    "ban on": "policy", "import ban": "policy", "export ban": "policy",
    "license": "policy", "licence": "policy", "compliance": "policy",
    "inflation": "policy", "devalu": "policy",
    # weather
    "flood": "weather", "rainfall": "weather", "drought": "weather",
    "heatwave": "weather", "harmattan": "weather", "storm": "weather",
    "nimet": "weather", "weather alert": "weather", "weather warning": "weather",
    "windstorm": "weather", "erosion": "weather",
    # infrastructure
    "power outage": "infrastructure", "blackout": "infrastructure",
    "electricity tariff": "infrastructure", "grid collapse": "infrastructure",
    "fuel scarcity": "infrastructure", "port congestion": "infrastructure",
    "road closure": "infrastructure", "bridge collaps": "infrastructure",
    "network outage": "infrastructure", "internet shutdown": "infrastructure",
    "telecom": "infrastructure", "logistics": "infrastructure",
    "supply chain": "infrastructure", "diesel price": "infrastructure",
    # funding / programs
    "grant": "funding", "loan scheme": "funding", "funding": "funding",
    "boi": "funding", "nirsan": "funding", "smedan": "funding",
    "empowerment program": "funding", "empowerment scheme": "funding",
    "accelerator": "funding", "incubat": "funding", "disburse": "funding",
    "credit facility": "funding", "intervention fund": "funding",
    "msme": "funding", "sme ": "funding", "startup fund": "funding",
    # security (business-operational, not generic crime)
    "kidnap": "security", "bandit": "security", "robbery": "security",
    "protest": "security", "strike action": "security", "shutdown order": "security",
    "curfew": "security", "unrest": "security", "insecurity": "security",
    "looting": "security", "market burnt": "security", "market fire": "security",
    # general business/economy signal not tied to one category
    "business": "other", "trade": "other", "commerce": "other", "market": "other",
    "manufactur": "other", "agribusiness": "other", "investment": "other",
    "economy": "other", "economic": "other", "entrepreneur": "other",
    "export": "other", "import": "other", "stock exchange": "other",
    "naira exchange": "other", "gdp": "other",
}

_PATTERN_OVERRIDES = {
    "son ": r"\bSON\b",
    "sme ": r"\bSMEs?\b",
}


def _compile(root: str) -> re.Pattern:
    if root in _PATTERN_OVERRIDES:
        return re.compile(_PATTERN_OVERRIDES[root], re.IGNORECASE)
    if " " in root:
        return re.compile(r"\b" + re.escape(root.strip()) + r"\w*\b", re.IGNORECASE)
    return re.compile(r"\b" + re.escape(root) + r"\w*\b", re.IGNORECASE)


_PATTERNS = {root: _compile(root) for root in BUSINESS_KEYWORDS}


def score(text: str) -> dict:
    """Cheap verdict on whether `text` is worth a Groq call.

    Zero matches -> not a candidate; classify_item() is skipped entirely and the
    item is auto-rejected. Any match -> candidate; Groq still does the real
    classification (category/impact/business_types/location), since a keyword
    hit only proves the text is business-adjacent, not which category it is.
    """
    matched = [root.strip() for root, pat in _PATTERNS.items() if pat.search(text)]
    is_candidate = bool(matched)
    reason = None if is_candidate else "no business-relevant keyword matched"
    return {
        "is_candidate": is_candidate,
        "matched_keywords": matched,
        "reason": reason,
    }