Blame for editor/answer.py

"""Answers questions from the site's own words. No model, no inference.

Everything returned here is a direct quote from published page content plus the
link it came from. If the site does not say it, this says so rather than
producing something plausible — a school site cannot afford invented answers
about pricing, dates or who is responsible for a child.
"""
import re

import mesh

STOPWORDS = {
    "a", "an", "the", "is", "are", "was", "were", "do", "does", "did", "i", "you",
    "my", "our", "we", "can", "could", "would", "should", "please", "tell", "me",
    "about", "for", "of", "to", "in", "on", "at", "and", "or", "it", "this", "that",
    "there", "here", "what", "whats", "when", "where", "who", "whos", "how", "why",
    "which", "have", "has", "any", "get", "find", "know", "want", "need", "am",
    "be", "with", "from", "if", "so", "up", "out", "will", "much", "many", "some",
}

# Question shape → how to phrase the reply. Order matters: first match wins.
INTENTS = [
    ("who",     re.compile(r"\bwho(?:'s|s\b|\s+is|\s+are|\s+teaches|\s+do)?\b", re.I),
     "Here is what the site says about who is involved:"),
    ("when",    re.compile(r"\bwhen\b|\bwhat time\b|\bdates?\b|\bschedule\b|\bcalendar\b|\bbreak\b|\bholiday\b", re.I),
     "Here are the dates and times listed:"),
    ("cost",    re.compile(r"\bcost|price|pricing|how much|fee|fees|pay|payment|free\b", re.I),
     "Here is the pricing information listed:"),
    ("where",   re.compile(r"\bwhere\b|\baddress\b|\blocation\b|\bdirections?\b", re.I),
     "Here is the location information listed:"),
    ("contact", re.compile(r"\bcontact\b|\bphone\b|\bemail\b|\bcall\b|\breach\b", re.I),
     "Here are the contact details listed:"),
    ("menu",    re.compile(r"\bmenu\b|\blunch\b|\bbreakfast\b|\bmeal\b|\bcafeteria\b", re.I),
     "Here is what the menu information says:"),
]
DEFAULT_LEAD = "Here is what the site says:"

DATE_RE = re.compile(
    r"\b(?:jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)[a-z]*\.?\s+\d{1,2}"
    r"|\b\d{1,2}/\d{1,2}(?:/\d{2,4})?"
    r"|\b(?:mon|tues|wednes|thurs|fri|satur|sun)day\b"
    r"|\b\d{1,2}:\d{2}\s*(?:a\.?m\.?|p\.?m\.?)?", re.I)
MONEY_RE = re.compile(r"\$\s?\d[\d,]*(?:\.\d{2})?")
PHONE_RE = re.compile(r"\b\(?\d{3}\)?[-.\s]\d{3}[-.\s]\d{4}\b")
EMAIL_RE = re.compile(r"\b[\w.+-]+@[\w-]+\.[\w.]+\b")

INTENT_PATTERNS = {"when": DATE_RE, "cost": MONEY_RE, "contact": PHONE_RE}


def keywords(question):
    words = re.findall(r"[\w']+", (question or "").lower())
    kept = [w for w in words if w not in STOPWORDS and len(w) > 2]
    return kept or words


def intent_of(question):
    for name, pattern, lead in INTENTS:
        if pattern.search(question or ""):
            return name, lead
    return "what", DEFAULT_LEAD


DANGLING = re.compile(r"\b(?:the|a|an|on|at|to|of|for|in|and|or|with|by|from|our|your)$", re.I)


def clean(text):
    """Editors write literal ** in this content; visitors should not see it."""
    text = re.sub(r"\*\*(.+?)\*\*", r"\1", text or "")
    text = text.replace("**", "").replace("__", "")
    return re.sub(r"\s+", " ", text).strip()


def sentences(text):
    parts = re.split(r"(?<=[.!?])\s+|\s*\n+\s*|\s{2,}", text or "")
    return [clean(s).strip(" \t·-–—") for s in parts if len(clean(s)) > 2]


def score_sentence(s, terms, intent):
    low = s.lower()
    score = sum(3 for t in terms if t in low)
    pattern = INTENT_PATTERNS.get(intent)
    if pattern and pattern.search(s):
        score += 4
    if intent == "contact" and (EMAIL_RE.search(s) or PHONE_RE.search(s)):
        score += 3
    # A bare heading like "Superintendent" repeats the question instead of
    # answering it. Reward sentences that actually carry information.
    if len(s) < 20:
        score -= 6
    elif len(s) >= 45:
        score += 2
    # "For school-specific contacts, visit the" — a fragment cut mid-thought
    # reads as broken. Prefer sentences that finish.
    if DANGLING.search(s.rstrip(" .,:;")):
        score -= 5
    score -= abs(len(s) - 130) / 400              # prefer a readable length
    return score


def best_sentence(text, terms, intent):
    """The one sentence that most answers the question, not the whole block."""
    best, best_score = "", -999
    for s in sentences(text):
        score = score_sentence(s, terms, intent)
        if score > best_score:
            best, best_score = s, score
    return best, best_score


def ask(question, limit=3):
    """Answer from published content only. Returns quotes plus where to read more."""
    terms = keywords(question)
    intent, lead = intent_of(question)
    if not terms:
        return {"ok": False, "intent": intent,
                "answer": "Ask me about something on the site — lunch menus, "
                          "calendars, staff, or a school office.",
                "sources": []}

    hits = mesh.public_search(" ".join(terms))
    if not hits:
        return {"ok": False, "intent": intent,
                "answer": f"I could not find anything on the site about "
                          f"“{question.strip()}”. Try different words, or use the "
                          f"site menu to browse.",
                "sources": []}

    # Collapse to the strongest passage per page.
    pages = {}
    for h in hits:
        aid = h["article_id"]
        quote, quality = best_sentence(h["text"], terms, intent)
        if not quote:
            continue
        # A headline names the topic; the body answers the question.
        if h["type"] == "headline":
            quality -= 4
        page = pages.get(aid)
        if page is None or quality > page["quality"]:
            pages[aid] = {"article_id": aid, "title": h["article"], "quote": quote,
                          "quality": quality, "coord": h["coord"],
                          "terms": page["terms"] if page else set()}
        pages[aid]["terms"] |= {t for t in terms if t in (h["text"] or "").lower()}

    # Rank pages by how much of the question they actually cover, not by any
    # single passage — a page mentioning every term is the likelier answer.
    ranked = sorted(pages.values(),
                    key=lambda p: (len(p["terms"]), p["quality"]), reverse=True)
    top = ranked[:limit]
    if not top:
        return {"ok": False, "intent": intent,
                "answer": "I found related pages but no clear sentence to quote. "
                          "Try opening one of them.",
                "sources": []}

    return {
        "ok": True,
        "intent": intent,
        "lead": lead,
        "answer": top[0]["quote"],
        "source": {"title": top[0]["title"], "article_id": top[0]["article_id"]},
        "sources": [{"title": t["title"], "article_id": t["article_id"], "quote": t["quote"]}
                    for t in top],
    }