from sentence_transformers import SentenceTransformer
import numpy as np
 
# 50 doc, her biri ~2K token
DOCS = [
    {"id": 1, "title": "Kargo Süresi", "content": "..."},
    {"id": 2, "title": "İade Politikası", "content": "..."},
    # ... 48 daha
]
 
# Türkçe embedding model
embedder = SentenceTransformer("sentence-transformers/paraphrase-multilingual-mpnet-base-v2")
 
doc_embeddings = embedder.encode([d["content"] for d in DOCS])
 
def retrieve_top_k(query, k=5):
    q_emb = embedder.encode([query])[0]
    scores = doc_embeddings @ q_emb
    top_idx = np.argsort(scores)[-k:][::-1]
    return [DOCS[i] for i in top_idx]

# A) Saf RAG
def pure_rag(query):
    docs = retrieve_top_k(query, k=5)
    context = "\n\n".join(d["content"] for d in docs)
    return client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=400,
        system=SHORT_SYSTEM,  # 500 token, cache yok
        messages=[{"role": "user", "content": f"{context}\n\nSoru: {query}"}],
    )
 
# B) Hybrid
HOT_TOP_10 = "\n\n".join(d["content"] for d in DOCS[:10])  # ilk 10 hot
 
def hybrid(query):
    docs = retrieve_top_k(query, k=5)
    # Hot olanları skip
    cold_docs = [d for d in docs if d["id"] > 10]
    cold_context = "\n\n".join(d["content"] for d in cold_docs)
 
    return client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=400,
        system=[
            {"type": "text", "text": SHORT_SYSTEM, "cache_control": {"type": "ephemeral"}},
            {"type": "text", "text": HOT_TOP_10, "cache_control": {"type": "ephemeral", "ttl": "1h"}},
        ],
        messages=[{"role": "user", "content": f"{cold_context}\n\nSoru: {query}"}],
    )

import time
from collections import defaultdict
 
QUERIES = [
    "Kargo süresi ne kadar?",
    "İade nasıl yapılır?",
    # ... 98 daha
]
 
stats = defaultdict(lambda: {"total_cost": 0.0, "total_latency": 0.0, "n": 0})
 
for name, fn in [("pure_rag", pure_rag), ("hybrid", hybrid)]:
    for q in QUERIES:
        start = time.perf_counter()
        resp = fn(q)
        latency = time.perf_counter() - start
 
        u = resp.usage
        cost = (
            u.input_tokens / 1e6 * 3.0
            + (getattr(u, "cache_creation_input_tokens", 0) or 0) / 1e6 * 3.75
            + (getattr(u, "cache_read_input_tokens", 0) or 0) / 1e6 * 0.30
            + u.output_tokens / 1e6 * 15.0
        )
        stats[name]["total_cost"] += cost
        stats[name]["total_latency"] += latency
        stats[name]["n"] += 1
 
# Sonuç
print(f"{'Mimari':<10} {'Cost':>10} {'TL':>10} {'Avg Latency':>14}")
print("─" * 50)
for name, s in stats.items():
    print(f"{name:<10} ${s['total_cost']:>8.4f} {s['total_cost']*33.5:>8.2f} {s['total_latency']/s['n']:>12.2f}s")