"""Prompt builder + safety guardrails + few-shots.

Distilled from step4_app.py to a much smaller, maintainable surface.
The 30-rule monolith was compressed to the rules that actually fire.
"""
from __future__ import annotations
from typing import List, Optional

# ── Safety guardrails: shown to the LLM, also enforced by citation_guard.py ──

SAFETY_GUARDRAILS = {
    "pest": (
        "SAFETY: Recommend ONLY ICAR-approved insecticides with EXACT dose and unit "
        "(e.g. 'Imidacloprid 17.8% SL @ 0.3 ml/L'). Never recommend banned chemicals "
        "(Endosulfan, Monocrotophos, Phorate, Phosphamidon)."
    ),
    "disease": (
        "SAFETY: Recommend ONLY ICAR-approved fungicides with EXACT dose and unit. "
        "Always include timing and method (foliar spray vs drench) and a 1-line IPM note."
    ),
    "nutrient": (
        "SAFETY: Quote standard ICAR fertiliser doses per acre/hectare. Mention soil "
        "test if the question is generic ('what fertiliser')."
    ),
    "weather": (
        "SAFETY: For frost/heat/drought/flood, give protective steps for the next 48-72h. "
        "Do not predict beyond Open-Meteo's 7-day horizon."
    ),
    "price": (
        "SAFETY: Do NOT quote prices from the retrieved KCC pairs (they may be years old). "
        "Direct the farmer to the live mandi forecast (P25/P50/P75) for current price ranges."
    ),
    "scheme": (
        "SAFETY: Quote scheme names and amounts ONLY if confident. Always end with the "
        "official portal URL (pmkisan.gov.in, pmfby.gov.in) so the farmer can verify."
    ),
}

# Hard rules that override retrieval if it suggests something dangerous.
HARD_OVERRIDES = [
    # (pest_or_disease_substring, banned_chemical_substring, replacement_advice)
    ("brown plant hopper", "imidacloprid",
     "For BPH (Brown Plant Hopper) on rice, NEVER use Imidacloprid (causes resurgence). "
     "Use Buprofezin 25% SC @ 1 ml/L or Pymetrozine 50% WG @ 0.6 g/L."),
    ("bph", "imidacloprid",
     "For BPH (Brown Plant Hopper) on rice, NEVER use Imidacloprid (causes resurgence). "
     "Use Buprofezin 25% SC @ 1 ml/L or Pymetrozine 50% WG @ 0.6 g/L."),
]

BANNED_CHEMICALS = {
    # India CIBRC banned + restricted (as of 2024)
    "endosulfan", "monocrotophos", "phorate", "phosphamidon",
    "methyl parathion", "carbosulfan", "phosalone",
    "aldrin", "heptachlor", "lindane", "ddt", "bhc",
    "chlordane", "dieldrin", "ethyl parathion", "fenthion",
    "tetradifon", "pentachlorophenol", "dicofol",
    "ethyl mercury chloride", "trichlorfon",
}

# Minimum citations required in answers for these problem types.
REQUIRES_CITATION = {"pest", "disease", "nutrient", "scheme"}


# ── Few-shot examples — problem-type filtered ────────────────────────────────

FEW_SHOTS = {
    "pest": [
        {
            "q": "My cotton has whitefly attack. What should I spray?",
            "a": ("**Diagnosis:** Whitefly (*Bemisia tabaci*) — common on cotton, causes "
                  "yellowing and sooty mould.\n"
                  "**Spray:** Diafenthiuron 50% WP @ 1.2 g/L of water [1] OR "
                  "Pyriproxyfen 10% EC @ 1 ml/L [2].\n"
                  "**Volume:** 200 L/acre. Spray under-leaf surface.\n"
                  "**Timing:** Evening, repeat after 10 days if needed.\n"
                  "**IPM:** Yellow sticky traps (5/acre); avoid Imidacloprid sprays in "
                  "boll-formation stage (resistance build-up).")
        },
    ],
    "disease": [
        {
            "q": "Wheat leaves have orange-red rust pustules. Treatment?",
            "a": ("**Diagnosis:** Wheat leaf/stripe rust (*Puccinia* spp.) — confirmed by "
                  "orange-red pustules.\n"
                  "**Spray:** Propiconazole 25% EC @ 1 ml/L [1] OR Tebuconazole 250 EC @ "
                  "1 ml/L [2].\n"
                  "**Volume:** 200 L/acre, ensure full canopy coverage.\n"
                  "**Timing:** At first sign of pustules; repeat after 12-15 days if rust "
                  "persists.\n"
                  "**IPM:** Sow rust-resistant varieties (HD-3086, DBW-187, PBW-725).")
        },
    ],
    "nutrient": [
        {
            "q": "Maize leaves are turning pale yellow. Which fertiliser?",
            "a": ("**Diagnosis:** Likely nitrogen deficiency (chlorosis spreads from older "
                  "leaves first).\n"
                  "**Top-dress:** Urea @ 50 kg/acre at knee-high stage [1] OR foliar Urea 2% "
                  "(20 g/L) for quick correction [2].\n"
                  "**Confirm:** Get a soil test — if pH > 8 the issue may be Zn or S, not N.")
        },
    ],
}


# ── Prompt builder ───────────────────────────────────────────────────────────

_SYSTEM = """You are an expert Indian agricultural advisor (KCC + ICAR). Answer farmer queries with practical, locally-relevant, ICAR-validated guidance.

ABSOLUTE RULES:
1. Give EXACT chemical doses with units (e.g. "1 ml/L" not "small amount").
2. NEVER recommend banned chemicals: Endosulfan, Monocrotophos, Phorate, Phosphamidon.
3. For pest/disease/nutrient/scheme answers, cite source numbers like [1] [2] from the context provided. If you cannot cite, say so explicitly.
4. Match the farmer's language: reply in Hindi if they wrote Hindi, English if English, Hinglish if Hinglish, etc.
5. Be concise (60-150 words). Structure: Diagnosis → Action → Dose → Timing → IPM/Safety.
6. If the retrieved Q&A do NOT support a confident answer, ask 1-2 clarifying questions instead of guessing.
"""


def build_prompt(query: str,
                 context: str,
                 *,
                 history: Optional[List[dict]] = None,
                 problem_type: str = "general",
                 language: str = "English",
                 detected_crop: Optional[str] = None,
                 state: str = "",
                 district: str = "",
                 low_confidence: bool = False) -> str:
    parts = [_SYSTEM]

    # Few-shot — pick at most 1 example matching the problem type.
    examples = FEW_SHOTS.get(problem_type) or []
    if examples:
        ex = examples[0]
        parts.append(f"\n### Example Q\n{ex['q']}\n### Example A\n{ex['a']}\n")

    # Safety note for this problem type
    safety = SAFETY_GUARDRAILS.get(problem_type, "")
    if safety:
        parts.append(f"\n[GUARDRAIL] {safety}")

    # Metadata banner
    meta = []
    if detected_crop:
        meta.append(f"Crop: {detected_crop}")
    if state:
        meta.append(f"State: {state}")
    if district:
        meta.append(f"District: {district}")
    if problem_type and problem_type != "general":
        meta.append(f"Problem type: {problem_type.upper()}")
    if meta:
        parts.append(f"\n[CONTEXT] {' | '.join(meta)}")

    if low_confidence:
        parts.append(
            "\n[LOW CONFIDENCE] Retrieved answers don't strongly match the query. "
            "Ask the farmer 1-2 clarifying questions BEFORE recommending a chemical."
        )

    # Conversation history (last 4 turns)
    if history:
        for turn in history[-4:]:
            role = turn.get("role", "user").capitalize()
            content = turn.get("content", "")
            if content:
                parts.append(f"\n{role}: {content}")

    parts.append(f"\n\n{context}")
    parts.append(f"\n\n### FARMER QUESTION (language: {language})\n{query}")
    parts.append(f"\n\n### YOUR ANSWER (in {language}, with citations [1][2] where applicable):")

    return "\n".join(parts)