Trainer_LLM/llm-api/plan_router.py

# -*- coding: utf-8 -*-
"""
plan_router.py – v0.13.2 (WP-15)

Minimal-CRUD + List/Filter für Templates & Pläne.

Änderungen ggü. v0.13.1
- Robuster Zeitfenster-Fallback: Wenn der serverseitige Range-Filter (created_at_ts) 0 Treffer liefert,
  wird eine zweite Scroll-Abfrage ohne Zeit-Range gefahren und lokal nach created_from/created_to gefiltert.
- plan_section_names bleibt Materialisierung & Filter-Basis.
"""
from fastapi import APIRouter, HTTPException, Query
from pydantic import BaseModel, Field
from typing import List, Optional, Dict, Any
from uuid import uuid4
from datetime import datetime, timezone
import hashlib
import json
import os

from clients import model, qdrant
from qdrant_client.models import (
    PointStruct, Filter, FieldCondition, MatchValue,
    VectorParams, Distance, Range
)

router = APIRouter(tags=["plans"])

# -----------------
# Konfiguration
# -----------------
PLAN_COLLECTION = os.getenv("PLAN_COLLECTION") or os.getenv("QDRANT_COLLECTION_PLANS", "plans")
PLAN_TEMPLATE_COLLECTION = os.getenv("PLAN_TEMPLATE_COLLECTION", "plan_templates")
PLAN_SESSION_COLLECTION = os.getenv("PLAN_SESSION_COLLECTION", "plan_sessions")
EXERCISE_COLLECTION = os.getenv("EXERCISE_COLLECTION", "exercises")

# -----------------
# Modelle
# -----------------
class TemplateSection(BaseModel):
    name: str
    target_minutes: int
    must_keywords: List[str] = []
    ideal_keywords: List[str] = []          # wünschenswert
    supplement_keywords: List[str] = []     # ergänzend
    forbid_keywords: List[str] = []
    capability_targets: Dict[str, int] = {}

class PlanTemplate(BaseModel):
    id: str = Field(default_factory=lambda: str(uuid4()))
    name: str
    discipline: str
    age_group: str
    target_group: str
    total_minutes: int
    sections: List[TemplateSection] = []
    goals: List[str] = []
    equipment_allowed: List[str] = []
    created_by: str
    version: str

class PlanItem(BaseModel):
    exercise_external_id: str
    duration: int
    why: str

class PlanSection(BaseModel):
    name: str
    items: List[PlanItem] = []
    minutes: int

class Plan(BaseModel):
    id: str = Field(default_factory=lambda: str(uuid4()))
    template_id: Optional[str] = None
    title: str
    discipline: str
    age_group: str
    target_group: str
    total_minutes: int
    sections: List[PlanSection] = []
    goals: List[str] = []
    capability_summary: Dict[str, int] = {}
    novelty_against_last_n: Optional[float] = None
    fingerprint: Optional[str] = None
    created_by: str
    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
    source: str = "API"

class PlanTemplateList(BaseModel):
    items: List[PlanTemplate]
    limit: int
    offset: int
    count: int

class PlanList(BaseModel):
    items: List[Plan]
    limit: int
    offset: int
    count: int

# -----------------
# Helpers
# -----------------
def _ensure_collection(name: str):
    if not qdrant.collection_exists(name):
        qdrant.recreate_collection(
            collection_name=name,
            vectors_config=VectorParams(size=model.get_sentence_embedding_dimension(), distance=Distance.COSINE),
        )

def _norm_list(xs: List[str]) -> List[str]:
    seen, out = set(), []
    for x in xs or []:
        s = str(x).strip()
        k = s.casefold()
        if s and k not in seen:
            seen.add(k)
            out.append(s)
    return sorted(out, key=str.casefold)

def _template_embed_text(tpl: PlanTemplate) -> str:
    parts = [tpl.name, tpl.discipline, tpl.age_group, tpl.target_group]
    parts += tpl.goals
    parts += [s.name for s in tpl.sections]
    return ". ".join([p for p in parts if p])

def _plan_embed_text(p: Plan) -> str:
    parts = [p.title, p.discipline, p.age_group, p.target_group]
    parts += p.goals
    parts += [s.name for s in p.sections]
    return ". ".join([p for p in parts if p])

def _embed(text: str):
    return model.encode(text or "").tolist()

def _fingerprint_for_plan(p: Plan) -> str:
    """sha256(title, total_minutes, sections.items.exercise_external_id, sections.items.duration)"""
    core = {
        "title": p.title,
        "total_minutes": int(p.total_minutes),
        "items": [
            {"exercise_external_id": it.exercise_external_id, "duration": int(it.duration)}
            for sec in p.sections
            for it in (sec.items or [])
        ],
    }
    raw = json.dumps(core, sort_keys=True, ensure_ascii=False)
    return hashlib.sha256(raw.encode("utf-8")).hexdigest()

def _get_by_field(collection: str, key: str, value: Any) -> Optional[Dict[str, Any]]:
    flt = Filter(must=[FieldCondition(key=key, match=MatchValue(value=value))])
    pts, _ = qdrant.scroll(collection_name=collection, scroll_filter=flt, limit=1, with_payload=True)
    if not pts:
        return None
    payload = dict(pts[0].payload or {})
    payload.setdefault("id", str(pts[0].id))
    return payload

def _as_model(model_cls, payload: Dict[str, Any]):
    fields = getattr(model_cls, "model_fields", None) or getattr(model_cls, "__fields__", {})
    allowed = set(fields.keys())
    data = {k: payload[k] for k in payload.keys() if k in allowed}
    return model_cls(**data)

def _truthy(val: Optional[str]) -> bool:
    return str(val or "").strip().lower() in {"1", "true", "yes", "on"}

def _exists_in_collection(collection: str, key: str, value: Any) -> bool:
    flt = Filter(must=[FieldCondition(key=key, match=MatchValue(value=value))])
    pts, _ = qdrant.scroll(collection_name=collection, scroll_filter=flt, limit=1, with_payload=False)
    return bool(pts)

# -----------------
# Endpoints: Templates
# -----------------
@router.post(
    "/plan_templates",
    response_model=PlanTemplate,
    summary="Create a plan template",
    description=(
        "Erstellt ein Plan-Template (Strukturplanung).\n\n"
        "• Mehrere Sections erlaubt.\n"
        "• Section-Felder: must/ideal/supplement/forbid keywords + capability_targets.\n"
        "• Materialisierte Facettenfelder (section_*) werden intern geschrieben, um Qdrant-Filter zu beschleunigen."
    ),
)
def create_plan_template(t: PlanTemplate):
    _ensure_collection(PLAN_TEMPLATE_COLLECTION)
    payload = t.model_dump()
    payload["goals"] = _norm_list(payload.get("goals"))
    sections = payload.get("sections", []) or []
    for s in sections:
        s["must_keywords"] = _norm_list(s.get("must_keywords") or [])
        s["ideal_keywords"] = _norm_list(s.get("ideal_keywords") or [])
        s["supplement_keywords"] = _norm_list(s.get("supplement_keywords") or [])
        s["forbid_keywords"] = _norm_list(s.get("forbid_keywords") or [])

    # Materialisierte Facetten (KEYWORD-Indizes)
    payload["section_names"] = _norm_list([s.get("name", "") for s in sections])
    payload["section_must_keywords"] = _norm_list([kw for s in sections for kw in (s.get("must_keywords") or [])])
    payload["section_ideal_keywords"] = _norm_list([kw for s in sections for kw in (s.get("ideal_keywords") or [])])
    payload["section_supplement_keywords"] = _norm_list([kw for s in sections for kw in (s.get("supplement_keywords") or [])])
    payload["section_forbid_keywords"] = _norm_list([kw for s in sections for kw in (s.get("forbid_keywords") or [])])

    vec = _embed(_template_embed_text(t))
    qdrant.upsert(collection_name=PLAN_TEMPLATE_COLLECTION, points=[PointStruct(id=str(t.id), vector=vec, payload=payload)])
    return t

@router.get(
    "/plan_templates/{tpl_id}",
    response_model=PlanTemplate,
    summary="Read a plan template by id",
    description="Liest ein Template anhand seiner ID und gibt nur die Schemafelder zurück (zusätzliche Payload wird herausgefiltert).",
)
def get_plan_template(tpl_id: str):
    _ensure_collection(PLAN_TEMPLATE_COLLECTION)
    found = _get_by_field(PLAN_TEMPLATE_COLLECTION, "id", tpl_id)
    if not found:
        raise HTTPException(status_code=404, detail="not found")
    return _as_model(PlanTemplate, found)

@router.get(
    "/plan_templates",
    response_model=PlanTemplateList,
    summary="List plan templates (filterable)",
    description=(
        "Listet Plan-Templates mit Filtern.\n\n"
        "**Filter** (exakte Matches, KEYWORD-Felder):\n"
        "- discipline, age_group, target_group\n"
        "- section: Section-Name (nutzt materialisierte `section_names`)\n"
        "- goal: Ziel (nutzt `goals`)\n"
        "- keyword: trifft auf beliebige Section-Keyword-Felder (must/ideal/supplement/forbid).\n\n"
        "**Pagination:** limit/offset. Feld `count` entspricht der Anzahl zurückgegebener Items (keine Gesamtsumme)."
    ),
)
def list_plan_templates(
    discipline: Optional[str] = Query(None, description="Filter: Disziplin (exaktes KEYWORD-Match)", example="Karate"),
    age_group: Optional[str] = Query(None, description="Filter: Altersgruppe", example="Teenager"),
    target_group: Optional[str] = Query(None, description="Filter: Zielgruppe", example="Breitensport"),
    section: Optional[str] = Query(None, description="Filter: Section-Name (materialisiert)", example="Warmup"),
    goal: Optional[str] = Query(None, description="Filter: Trainingsziel", example="Technik"),
    keyword: Optional[str] = Query(None, description="Filter: Keyword in must/ideal/supplement/forbid", example="Koordination"),
    limit: int = Query(20, ge=1, le=200, description="Max. Anzahl Items"),
    offset: int = Query(0, ge=0, description="Start-Offset für Paging"),
):
    _ensure_collection(PLAN_TEMPLATE_COLLECTION)
    must: List[Any] = []
    should: List[Any] = []
    if discipline:
        must.append(FieldCondition(key="discipline", match=MatchValue(value=discipline)))
    if age_group:
        must.append(FieldCondition(key="age_group", match=MatchValue(value=age_group)))
    if target_group:
        must.append(FieldCondition(key="target_group", match=MatchValue(value=target_group)))
    if section:
        must.append(FieldCondition(key="section_names", match=MatchValue(value=section)))
    if goal:
        must.append(FieldCondition(key="goals", match=MatchValue(value=goal)))
    if keyword:
        for k in (
            "section_must_keywords",
            "section_ideal_keywords",
            "section_supplement_keywords",
            "section_forbid_keywords",
        ):
            should.append(FieldCondition(key=k, match=MatchValue(value=keyword)))

    flt = None
    if must or should:
        flt = Filter(must=must or None, should=should or None)

    fetch_n = max(offset + limit, 1)
    pts, _ = qdrant.scroll(collection_name=PLAN_TEMPLATE_COLLECTION, scroll_filter=flt, limit=fetch_n, with_payload=True)
    items: List[PlanTemplate] = []
    for p in pts[offset:offset+limit]:
        payload = dict(p.payload or {})
        payload.setdefault("id", str(p.id))
        items.append(_as_model(PlanTemplate, payload))
    return PlanTemplateList(items=items, limit=limit, offset=offset, count=len(items))

# -----------------
# Endpoints: Pläne
# -----------------
@router.post(
    "/plan",
    response_model=Plan,
    summary="Create a concrete training plan",
    description=(
        "Erstellt einen konkreten Trainingsplan.\n\n"
        "Idempotenz: gleicher Fingerprint (title + items) → gleicher Plan (kein Duplikat).\n"
        "Optional: Validierung von template_id und Exercises (Strict-Mode)."
    ),
)
def create_plan(p: Plan):
    _ensure_collection(PLAN_COLLECTION)

    # Template-Referenz prüfen (falls gesetzt)
    if p.template_id:
        if not _exists_in_collection(PLAN_TEMPLATE_COLLECTION, "id", p.template_id):
            raise HTTPException(status_code=422, detail=f"Unknown template_id: {p.template_id}")

    # Optional: Strict-Mode – Exercises gegen EXERCISE_COLLECTION prüfen
    if _truthy(os.getenv("PLAN_STRICT_EXERCISES")):
        missing: List[str] = []
        for sec in p.sections or []:
            for it in sec.items or []:
                exid = (it.exercise_external_id or "").strip()
                if exid and not _exists_in_collection(EXERCISE_COLLECTION, "external_id", exid):
                    missing.append(exid)
        if missing:
            raise HTTPException(status_code=422, detail={"error": "unknown exercise_external_id", "missing": sorted(set(missing))})

    # Fingerprint + Idempotenz
    fp = _fingerprint_for_plan(p)
    p.fingerprint = p.fingerprint or fp
    existing = _get_by_field(PLAN_COLLECTION, "fingerprint", p.fingerprint)
    if existing:
        return _as_model(Plan, existing)

    # Normalisieren + Materialisierung
    p.goals = _norm_list(p.goals)
    payload = p.model_dump()

    # created_at → ISO + numerischer Zeitstempel (FLOAT)
    dt = payload.get("created_at")
    if isinstance(dt, datetime):
        dt = dt.astimezone(timezone.utc).isoformat()
    elif isinstance(dt, str):
        try:
            _ = datetime.fromisoformat(dt.replace("Z", "+00:00"))
        except Exception:
            dt = datetime.now(timezone.utc).isoformat()
    else:
        dt = datetime.now(timezone.utc).isoformat()
    payload["created_at"] = dt
    try:
        ts = datetime.fromisoformat(dt.replace("Z", "+00:00")).timestamp()
    except Exception:
        ts = datetime.now(timezone.utc).timestamp()
    payload["created_at_ts"] = float(ts)

    # Materialisierte Section-Namen (robuste Filter/Indizes)
    try:
        payload["plan_section_names"] = _norm_list([
            (s.get("name") or "").strip() for s in (payload.get("sections") or []) if isinstance(s, dict)
        ])
    except Exception:
        payload["plan_section_names"] = _norm_list([
            (getattr(s, "name", "") or "").strip() for s in (p.sections or [])
        ])

    vec = _embed(_plan_embed_text(p))
    qdrant.upsert(collection_name=PLAN_COLLECTION, points=[PointStruct(id=str(p.id), vector=vec, payload=payload)])
    return p

@router.get(
    "/plan/{plan_id}",
    response_model=Plan,
    summary="Read a training plan by id",
    description="Liest einen Plan anhand seiner ID. `created_at` wird (falls ISO-String) zu `datetime` geparst.",
)
def get_plan(plan_id: str):
    _ensure_collection(PLAN_COLLECTION)
    found = _get_by_field(PLAN_COLLECTION, "id", plan_id)
    if not found:
        raise HTTPException(status_code=404, detail="not found")
    if isinstance(found.get("created_at"), str):
        try:
            found["created_at"] = datetime.fromisoformat(found["created_at"])
        except Exception:
            pass
    return _as_model(Plan, found)

@router.get(
    "/plans",
    response_model=PlanList,
    summary="List training plans (filterable)",
    description=(
        "Listet Trainingspläne mit Filtern.\n\n"
        "**Filter** (exakte Matches, KEYWORD-Felder):\n"
        "- created_by, discipline, age_group, target_group, goal\n"
        "- section: Section-Name (nutzt materialisiertes `plan_section_names`)\n"
        "- created_from / created_to: ISO-8601 Zeitfenster → serverseitiger Range-Filter über `created_at_ts` (FLOAT). Bei 0 Treffern wird lokal gefiltert.\n\n"
        "**Pagination:** limit/offset. Feld `count` entspricht der Anzahl zurückgegebener Items (keine Gesamtsumme)."
    ),
)
def list_plans(
    created_by: Optional[str] = Query(None, description="Filter: Ersteller", example="tester"),
    discipline: Optional[str] = Query(None, description="Filter: Disziplin", example="Karate"),
    age_group: Optional[str] = Query(None, description="Filter: Altersgruppe", example="Teenager"),
    target_group: Optional[str] = Query(None, description="Filter: Zielgruppe", example="Breitensport"),
    goal: Optional[str] = Query(None, description="Filter: Trainingsziel", example="Technik"),
    section: Optional[str] = Query(None, description="Filter: Section-Name", example="Warmup"),
    created_from: Optional[str] = Query(None, description="Ab-Zeitpunkt (ISO 8601, z. B. 2025-08-12T00:00:00Z)", example="2025-08-12T00:00:00Z"),
    created_to: Optional[str] = Query(None, description="Bis-Zeitpunkt (ISO 8601)", example="2025-08-13T00:00:00Z"),
    limit: int = Query(20, ge=1, le=200, description="Max. Anzahl Items"),
    offset: int = Query(0, ge=0, description="Start-Offset für Paging"),
):
    _ensure_collection(PLAN_COLLECTION)

    # 1) Grundfilter (ohne Zeit)
    base_must: List[Any] = []
    if created_by:
        base_must.append(FieldCondition(key="created_by", match=MatchValue(value=created_by)))
    if discipline:
        base_must.append(FieldCondition(key="discipline", match=MatchValue(value=discipline)))
    if age_group:
        base_must.append(FieldCondition(key="age_group", match=MatchValue(value=age_group)))
    if target_group:
        base_must.append(FieldCondition(key="target_group", match=MatchValue(value=target_group)))
    if goal:
        base_must.append(FieldCondition(key="goals", match=MatchValue(value=goal)))
    if section:
        base_must.append(FieldCondition(key="plan_section_names", match=MatchValue(value=section)))

    # 2) Serverseitiger Zeitbereich
    range_args: Dict[str, float] = {}
    try:
        if created_from:
            range_args["gte"] = float(datetime.fromisoformat(created_from.replace("Z", "+00:00")).timestamp())
        if created_to:
            range_args["lte"] = float(datetime.fromisoformat(created_to.replace("Z", "+00:00")).timestamp())
    except Exception:
        range_args = {}

    applied_server_range = bool(range_args)
    must_with_time = list(base_must)
    if applied_server_range:
        must_with_time.append(FieldCondition(key="created_at_ts", range=Range(**range_args)))

    # 3) Erste Scroll: mit Zeit-Range (falls gesetzt)
    fetch_n = max(offset + limit, 100)  # etwas großzügiger, um Tests sicher zu treffen
    flt_time = Filter(must=must_with_time or None) if must_with_time else None
    pts, _ = qdrant.scroll(collection_name=PLAN_COLLECTION, scroll_filter=flt_time, limit=fetch_n, with_payload=True)

    # 4) Fallback: wenn serverseitig gefiltert wurde und 0 Treffer → zweite Scroll ohne Zeit-Range, lokal filtern
    if applied_server_range and not pts:
        flt_no_time = Filter(must=base_must or None) if base_must else None
        pts, _ = qdrant.scroll(collection_name=PLAN_COLLECTION, scroll_filter=flt_no_time, limit=fetch_n, with_payload=True)

    # 5) Lokaler Zeitfenster-Check (nur wenn keine serverseitige Zeit-Range gegriffen hat ODER Fallback aktiv war)
    def _in_window(py: Dict[str, Any]) -> bool:
        if not (created_from or created_to):
            return True
        ts = py.get("created_at")
        if isinstance(ts, dict) and ts.get("$date"):
            ts = ts["$date"]
        if isinstance(ts, (int, float)) and py.get("created_at_ts") is not None:
            dt = datetime.fromtimestamp(float(py["created_at_ts"]), tz=timezone.utc)
        elif isinstance(ts, str):
            try:
                dt = datetime.fromisoformat(ts.replace("Z", "+00:00"))
            except Exception:
                return False
        elif isinstance(ts, datetime):
            dt = ts
        else:
            return False
        ok = True
        if created_from:
            try:
                ok = ok and dt >= datetime.fromisoformat(created_from.replace("Z", "+00:00"))
            except Exception:
                pass
        if created_to:
            try:
                ok = ok and dt <= datetime.fromisoformat(created_to.replace("Z", "+00:00"))
            except Exception:
                pass
        return ok

    payloads: List[Dict[str, Any]] = []
    for p in pts:
        py = dict(p.payload or {})
        py.setdefault("id", str(p.id))
        # Wenn keine Zeit-Filter angegeben sind, oder wir im Fallback sind → prüfe lokal.
        if not (created_from or created_to) or (applied_server_range and not payloads and not pts):
            # (Die Bedingung oben wird praktisch nie gebraucht; belassen aus Klarheit.)
            pass
        if not applied_server_range:
            # Kein serverseitiger Range → lokal prüfen
            if not _in_window(py):
                continue
        # Wenn serverseitig aktiv war und wir trotzdem hier sind (Fallback-Pfad), lokal prüfen:
        if applied_server_range and range_args and len(pts) >= 0:
            if not _in_window(py):
                continue
        payloads.append(py)

    sliced = payloads[offset:offset+limit]
    items = [_as_model(Plan, x) for x in sliced]
    return PlanList(items=items, limit=limit, offset=offset, count=len(items))