"""
Live Accuracy and Usage Benchmark for OpenRouter BGE-M3 + Cohere Rerank
Sahih Islam Production Verification
"""
import sys
import os
import asyncio
import time
import numpy as np

# Ensure UTF-8 output on Windows
if sys.platform == "win32":
    sys.stdout.reconfigure(encoding="utf-8")

backend_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if backend_dir not in sys.path:
    sys.path.insert(0, backend_dir)

from dotenv import load_dotenv
load_dotenv(os.path.join(backend_dir, ".env"))

from app.database import hadith_collection
from app.inference_client import get_embedding, get_rerank_scores, DEFAULT_EMBEDDING_MODEL, DEFAULT_RERANKER_MODEL

# 5 Core Islamic Test Queries with Verified Golden Targets
BENCHMARK_QUERIES = [
    {
        "id": 1,
        "query": "actions are judged by intentions",
        "topic": "Intentions (Niyyah)",
        "expected_keywords": ["intention", "niyyah", "emigrate", "reward", "action"],
        "famous_ref": "Sahih Bukhari Hadith 1 / Muslim 1907"
    },
    {
        "id": 2,
        "query": "Islam is built upon five pillars",
        "topic": "Pillars of Islam (Arkan al-Islam)",
        "expected_keywords": ["pillar", "five", "prayer", "zakat", "fast", "hajj", "allah"],
        "famous_ref": "Sahih Bukhari Hadith 8 / Muslim 16"
    },
    {
        "id": 3,
        "query": "smiling at your brother is charity",
        "topic": "Charity in Good Deeds",
        "expected_keywords": ["smile", "charity", "brother", "sadaqah", "good"],
        "famous_ref": "Jami at-Tirmidhi 1956"
    },
    {
        "id": 4,
        "query": "the strong person is not the good wrestler but the one who controls anger",
        "topic": "Controlling Anger",
        "expected_keywords": ["wrestler", "strong", "anger", "angry", "controls"],
        "famous_ref": "Sahih Bukhari 6114 / Muslim 2609"
    },
    {
        "id": 5,
        "query": "who is most deserving of my good companionship mother",
        "topic": "Rights of Parents (Mother)",
        "expected_keywords": ["mother", "father", "companionship", "treatment", "good"],
        "famous_ref": "Sahih Bukhari 5971 / Muslim 2548"
    }
]

async def run_hadith_search(query: str, top_k: int = 3):
    t0 = time.time()
    
    # STEP 1: Bi-Encoder Dense Vectorization via OpenRouter BGE-M3
    query_vec = (await get_embedding(query)).tolist()
    t_embed = time.time() - t0

    # STEP 2: MongoDB Atlas Vector Search
    t1 = time.time()
    stage = {
        "$vectorSearch": {
            "index": "vector_index",
            "path": "embedding_1024",
            "queryVector": query_vec,
            "numCandidates": 200,
            "limit": 20
        }
    }
    raw_candidates = await hadith_collection.aggregate([
        stage,
        {"$addFields": {"_vscore": {"$meta": "vectorSearchScore"}}}
    ]).to_list(length=20)
    t_atlas = time.time() - t1

    if not raw_candidates:
        return [], t_embed, t_atlas, 0.0

    # STEP 3: Cross-Encoder Precision Reranking via OpenRouter Cohere Rerank v3.5
    t2 = time.time()
    pairs = [[query, str(d.get("text_english", ""))[:1000]] for d in raw_candidates]
    scores = await get_rerank_scores(pairs)
    t_rerank = time.time() - t2

    ranked_indices = np.argsort(scores)[::-1]
    
    final_results = []
    for idx in ranked_indices[:top_k]:
        doc = raw_candidates[idx]
        doc["_rerank_score"] = float(scores[idx])
        final_results.append(doc)

    return final_results, t_embed, t_atlas, t_rerank

async def main():
    print("=" * 80)
    print("SAHIH ISLAM - LIVE SEARCH ACCURACY BENCHMARK (OpenRouter Cloud Inference)")
    print(f"Bi-Encoder Model : {DEFAULT_EMBEDDING_MODEL}")
    print(f"Cross-Encoder    : {DEFAULT_RERANKER_MODEL}")
    print(f"Total Test Runs  : {len(BENCHMARK_QUERIES)} Searches")
    print("=" * 80)

    total_accuracy_score = 0
    total_searches = len(BENCHMARK_QUERIES)

    for item in BENCHMARK_QUERIES:
        q_id = item["id"]
        query = item["query"]
        print(f"\n[SEARCH #{q_id}/{total_searches}] Query: \"{query}\"")
        print(f"   Topic: {item['topic']} | Expected: {item['famous_ref']}")

        results, t_emb, t_atl, t_rer = await run_hadith_search(query, top_k=3)
        print(f"   Timing: Embed={t_emb:.3f}s | Atlas={t_atl:.3f}s | Rerank={t_rer:.3f}s | Total={t_emb+t_atl+t_rer:.3f}s")

        if not results:
            print("   ❌ ZERO RESULTS RETURNED!")
            continue

        top_hit = results[0]
        ref = top_hit.get("reference") or f"{top_hit.get('collection')} #{top_hit.get('hadith_no')}"
        narrator = top_hit.get("narrator", "N/A")
        text_en = top_hit.get("text_english", "")[:280]
        vscore = top_hit.get("_vscore", 0.0)
        rscore = top_hit.get("_rerank_score", 0.0)

        # Check keyword presence in top 3 hits
        found_keywords = [kw for kw in item["expected_keywords"] if kw in text_en.lower()]
        accuracy_match = len(found_keywords) >= 2

        if accuracy_match:
            total_accuracy_score += 1
            status_icon = "[EXACT MATCH]"
        else:
            status_icon = "[PARTIAL MATCH]"

        print(f"   {status_icon} -> Top Rank #1:")
        print(f"      Reference   : {ref}")
        print(f"      Narrator    : {narrator}")
        print(f"      Vector Sim  : {vscore:.4f} | Rerank Score: {rscore:.4f}")
        print(f"      Text Preview: \"{text_en}...\"")

    print("\n" + "=" * 80)
    print("BENCHMARK SUMMARY")
    print("=" * 80)
    print(f"Total Searches Executed        : {total_searches}")
    print(f"Total OpenRouter BGE-M3 Calls  : {total_searches}")
    print(f"Total OpenRouter Rerank Calls  : {total_searches}")
    print(f"Search Accuracy Rate           : {total_accuracy_score}/{total_searches} ({(total_accuracy_score/total_searches)*100:.1f}%)")
    print("=" * 80)

if __name__ == "__main__":
    asyncio.run(main())