src/pipeline/support_plan.py

"""Support Plan / clinician-handoff export.

Renders a Markdown summary of a conversation that the student can save and
optionally bring to a UMD CAPS counselor, ISSS advisor, ADS coordinator, or
graduate Ombuds. The plan is the student's record, never auto-shared.

Sections:

* Header — timestamp, anonymized session id, scope disclaimer
* What I'm working on — paraphrase of the user's first message + later messages
* What I've tried — placeholder; user fills in by hand
* What the navigator surfaced — routes/tiers/recommended action across turns
* Resources mentioned — deduped list with URLs

Tone is plain and stigma-free; never uses clinical labels and never claims to
have diagnosed or treated anything. The deterministic planner already enforces
this; the export simply preserves it.
"""

from __future__ import annotations

from dataclasses import dataclass
from datetime import datetime
from typing import Iterable


@dataclass
class TurnLogEntry:
    turn_index: int
    timestamp: str
    user_message: str
    route_label: str
    safety_tier: str
    conversation_stage: str
    recommended_action: str
    international_concern: bool
    intl_topic: str
    retrieved_sources: list[dict]


_PRETTY_ROUTE = {
    "academic_setback": "Academic setback",
    "exam_stress": "Exam or test stress",
    "accessibility_ads": "Accessibility / accommodations",
    "advisor_conflict": "Advisor or graduate conflict",
    "counseling_navigation": "Counseling navigation",
    "basic_needs": "Basic needs",
    "care_violence_confidential": "Confidential CARE support",
    "peer_helper": "Helping someone else",
    "loneliness_isolation": "Loneliness or isolation",
    "anxiety_panic": "Anxiety or panic",
    "low_mood": "Low mood",
    "crisis_immediate": "Immediate safety handoff",
    "general_student_support": "General student support",
    "out_of_scope": "Outside support scope",
}


def build_support_plan_markdown(
    turn_log: Iterable[dict | TurnLogEntry],
    started_at: datetime | None = None,
) -> str:
    """Build the Markdown support-plan body.

    Accepts dicts (the demo's session-state shape) or TurnLogEntry objects.
    """
    turns: list[dict] = [_normalize(t) for t in turn_log]
    if not turns:
        return _empty_plan(started_at)

    started = (started_at or datetime.utcnow()).strftime("%Y-%m-%d %H:%M UTC")

    user_msgs = [t["user_message"] for t in turns if t.get("user_message")]
    routes = _dedupe_in_order([t["route_label"] for t in turns if t.get("route_label")])
    intl_topics = _dedupe_in_order([t["intl_topic"] for t in turns if t.get("intl_topic")])
    actions = _dedupe_in_order([t["recommended_action"] for t in turns if t.get("recommended_action")])
    resources = _dedupe_resources(turns)

    lines: list[str] = []
    lines.append("# Support plan")
    lines.append("")
    lines.append(f"_Generated by EmpathRAG, a UMD support-navigation prototype._")
    lines.append(f"_Created: {started}._")
    lines.append("")
    lines.append("> This is your record. EmpathRAG is not therapy, diagnosis, or emergency care.")
    lines.append("> The navigator only points to UMD resources; it does not interpret policy on your behalf.")
    lines.append("")

    lines.append("## What I'm working on")
    lines.append("")
    if user_msgs:
        lines.append("In my own words:")
        lines.append("")
        for m in user_msgs:
            lines.append(f"- {m.strip()}")
    else:
        lines.append("_No messages recorded yet._")
    lines.append("")

    lines.append("## What I've tried")
    lines.append("")
    lines.append("- _Fill in by hand: things I've already attempted, who I've already talked to, what helped or didn't._")
    lines.append("")

    lines.append("## What the navigator surfaced")
    lines.append("")
    if routes:
        pretty = ", ".join(_PRETTY_ROUTE.get(r, r.replace("_", " ").title()) for r in routes)
        lines.append(f"- Topics identified: {pretty}")
    if intl_topics:
        pretty = ", ".join(t.replace("_", " ") for t in intl_topics)
        lines.append(f"- F-1 / international sub-topics: {pretty}")
    if actions:
        lines.append("- Suggested next steps:")
        for a in actions:
            lines.append(f"  - {a.strip().rstrip('.')}")
    lines.append("")

    lines.append("## Resources mentioned")
    lines.append("")
    if resources:
        for r in resources:
            line = f"- **{r['source_name']}**"
            if r.get("title"):
                line += f" — {r['title']}"
            if r.get("url") and r["url"] not in {"", "N/A"}:
                line += f" ({r['url']})"
            lines.append(line)
    else:
        lines.append("_None surfaced yet — keep talking to see relevant resources._")
    lines.append("")

    lines.append("---")
    lines.append("")
    lines.append("_If you bring this to a counselor or ISSS advisor: feel free to paste their notes below._")
    lines.append("")
    return "\n".join(lines)


def _locate_unicode_font() -> tuple[str | None, str | None]:
    """Return (regular_ttf, bold_ttf) paths for a Unicode-capable font, or
    (None, None) if we can't find one. We try DejaVu Sans (bundled with
    matplotlib in this venv), then DejaVu Sans installed system-wide, then
    Arial (Windows). Returns None paths if nothing usable is found; the PDF
    builder then degrades to the built-in Helvetica (latin-1 only)."""
    from pathlib import Path as _P
    import sys as _sys
    # matplotlib bundles DejaVu Sans in this venv; it covers Latin, Cyrillic,
    # Greek, IPA, common accented characters. Doesn't cover CJK but is
    # sufficient for the common case (accented Western names).
    try:
        import matplotlib  # type: ignore
        mpl_fonts = _P(matplotlib.__file__).parent / "mpl-data" / "fonts" / "ttf"
        reg = mpl_fonts / "DejaVuSans.ttf"
        bold = mpl_fonts / "DejaVuSans-Bold.ttf"
        if reg.exists() and bold.exists():
            return str(reg), str(bold)
    except Exception:
        pass
    # Windows fallback
    if _sys.platform.startswith("win"):
        win_fonts = _P("C:/Windows/Fonts")
        for reg_name, bold_name in [("arial.ttf", "arialbd.ttf"), ("calibri.ttf", "calibrib.ttf")]:
            reg = win_fonts / reg_name
            bold = win_fonts / bold_name
            if reg.exists() and bold.exists():
                return str(reg), str(bold)
    return None, None


def build_support_plan_pdf(
    turn_log: Iterable[dict | TurnLogEntry],
    out_path: str,
    started_at: datetime | None = None,
) -> str:
    """Render the support plan as a counselor-friendly PDF.

    Uses fpdf2 (small, pure-Python). Loads a Unicode-capable TTF (DejaVu Sans,
    bundled with matplotlib) so accented characters in student names —
    "José", "Müller", "李" — render correctly instead of being substituted
    with "?". Falls back to Helvetica + latin-1 encode if no Unicode font is
    available; that path still works for ASCII names but garbles accents.
    """
    from fpdf import FPDF

    turns = [_normalize(t) for t in turn_log]
    started = (started_at or datetime.utcnow()).strftime("%Y-%m-%d %H:%M UTC")

    pdf = FPDF(orientation="P", unit="mm", format="A4")
    pdf.set_auto_page_break(auto=True, margin=18)
    pdf.add_page()
    pdf.set_margins(left=18, top=18, right=18)

    # Try to register a Unicode font. If it works, use "Body" / "Body-B" as
    # font names below. If it fails, fall back to built-in Helvetica.
    reg_path, bold_path = _locate_unicode_font()
    use_unicode = False
    if reg_path and bold_path:
        try:
            pdf.add_font("Body", "", reg_path, uni=True)
            pdf.add_font("Body", "B", bold_path, uni=True)
            pdf.add_font("Body", "I", reg_path, uni=True)   # italic falls back to regular
            use_unicode = True
        except Exception:
            use_unicode = False
    family = "Body" if use_unicode else "Helvetica"

    def _safe(text: str) -> str:
        """If we're stuck with Helvetica (latin-1 only), strip characters
        outside latin-1 so fpdf2 doesn't crash. With the Unicode font,
        pass-through unchanged."""
        if use_unicode:
            return text
        return text.encode("latin-1", errors="replace").decode("latin-1")

    # Title block
    pdf.set_font(family, "B", 18)
    pdf.set_text_color(15, 23, 42)
    pdf.cell(0, 9, "Support plan", ln=True)
    pdf.set_font(family, "", 10)
    pdf.set_text_color(90, 100, 110)
    pdf.cell(0, 5, "Generated by EmpathRAG, a UMD support-navigation prototype.", ln=True)
    pdf.cell(0, 5, f"Created: {started}.", ln=True)
    pdf.ln(4)

    # Scope disclaimer box
    pdf.set_fill_color(240, 250, 248)
    pdf.set_draw_color(94, 234, 212)
    pdf.set_text_color(20, 60, 50)
    pdf.set_font(family, "", 10)
    pdf.multi_cell(
        0, 5,
        "This is your record. EmpathRAG is not therapy, diagnosis, or "
        "emergency care. The navigator only points to UMD resources; it does "
        "not interpret policy on your behalf.",
        border=1, fill=True,
    )
    pdf.ln(4)
    pdf.set_text_color(15, 23, 42)

    def section(title: str) -> None:
        pdf.set_font(family, "B", 12)
        pdf.set_text_color(20, 130, 110)
        pdf.cell(0, 7, title, ln=True)
        pdf.set_font(family, "", 11)
        pdf.set_text_color(15, 23, 42)

    def bullet(text: str) -> None:
        pdf.set_x(22)
        # ASCII bullet (fpdf2's default helvetica covers ASCII reliably; we
        # strip any chars outside latin-1 to avoid encoding errors).
        pdf.multi_cell(0, 5.5, f"- {_safe(text)}")

    # What I'm working on
    section("What I'm working on")
    user_msgs = [t["user_message"] for t in turns if t.get("user_message")]
    if user_msgs:
        pdf.set_font(family, "", 11)
        pdf.cell(0, 5.5, "In my own words:", ln=True)
        pdf.ln(1)
        for m in user_msgs:
            bullet(m.strip())
    else:
        pdf.set_font(family, "I", 11)
        pdf.set_text_color(120, 130, 140)
        pdf.cell(0, 5.5, "No messages recorded yet.", ln=True)
        pdf.set_text_color(15, 23, 42)
    pdf.ln(3)

    # What I've tried
    section("What I've tried")
    pdf.set_font(family, "I", 11)
    pdf.set_text_color(120, 130, 140)
    pdf.multi_cell(
        0, 5.5,
        "Fill in by hand: things I've already attempted, who I've already "
        "talked to, what helped or didn't.",
    )
    pdf.set_text_color(15, 23, 42)
    pdf.ln(3)

    # What the navigator surfaced
    section("What the navigator surfaced")
    routes = _dedupe_in_order([t["route_label"] for t in turns if t.get("route_label")])
    intl_topics = _dedupe_in_order([t["intl_topic"] for t in turns if t.get("intl_topic")])
    actions = _dedupe_in_order([t["recommended_action"] for t in turns if t.get("recommended_action")])
    if routes:
        pretty = ", ".join(_PRETTY_ROUTE.get(r, r.replace("_", " ").title()) for r in routes)
        bullet(f"Topics identified: {pretty}")
    if intl_topics:
        pretty = ", ".join(t.replace("_", " ") for t in intl_topics)
        bullet(f"F-1 / international sub-topics: {pretty}")
    if actions:
        pdf.set_x(22)
        pdf.set_font(family, "", 11)
        pdf.cell(0, 5.5, "- Suggested next steps:", ln=True)
        for a in actions:
            pdf.set_x(28)
            pdf.multi_cell(0, 5.5, f"  - {_safe(a.strip().rstrip('.'))}")
    pdf.ln(3)

    # Resources mentioned
    section("Resources mentioned")
    resources = _dedupe_resources(turns)
    if resources:
        for r in resources:
            line = r["source_name"]
            if r.get("title"):
                line += f" - {r['title']}"
            url = r.get("url") or ""
            if url and url not in {"", "N/A"}:
                line += f"  ({url})"
            bullet(line)
    else:
        pdf.set_font(family, "I", 11)
        pdf.set_text_color(120, 130, 140)
        pdf.cell(0, 5.5, "None surfaced yet - keep talking to see relevant resources.", ln=True)
        pdf.set_text_color(15, 23, 42)
    pdf.ln(6)

    # Footer / counselor handoff line
    pdf.set_font(family, "I", 9)
    pdf.set_text_color(120, 130, 140)
    pdf.multi_cell(
        0, 4.5,
        "If you bring this to a counselor or ISSS advisor: feel free to write "
        "their notes alongside.",
    )

    pdf.output(out_path)
    return out_path


def _empty_plan(started_at: datetime | None) -> str:
    started = (started_at or datetime.utcnow()).strftime("%Y-%m-%d %H:%M UTC")
    return (
        "# Support plan\n\n"
        f"_Generated: {started}._\n\n"
        "_No conversation yet. Send a message first, then download._\n"
    )


def _normalize(t: dict | TurnLogEntry) -> dict:
    if isinstance(t, TurnLogEntry):
        return {
            "turn_index": t.turn_index,
            "timestamp": t.timestamp,
            "user_message": t.user_message,
            "route_label": t.route_label,
            "safety_tier": t.safety_tier,
            "conversation_stage": t.conversation_stage,
            "recommended_action": t.recommended_action,
            "international_concern": t.international_concern,
            "intl_topic": t.intl_topic,
            "retrieved_sources": t.retrieved_sources,
        }
    return dict(t)


def _dedupe_in_order(items: list[str]) -> list[str]:
    seen: set[str] = set()
    out: list[str] = []
    for x in items:
        if not x:
            continue
        if x in seen:
            continue
        seen.add(x)
        out.append(x)
    return out


def _dedupe_resources(turns: list[dict]) -> list[dict]:
    """Collapse retrieved_sources across turns by (source_name, title)."""
    seen: set[tuple[str, str]] = set()
    out: list[dict] = []
    for t in turns:
        for r in t.get("retrieved_sources", []) or []:
            key = (r.get("source_name", ""), r.get("title", ""))
            if not key[0]:
                continue
            if key in seen:
                continue
            seen.add(key)
            out.append(r)
    return out