""" Benchmark / Threshold Analizi ----------------------------------- 30 test sorusunu (eval/test_questions.json) çalıştırır: - Her soru için embed edilir, ChromaDB'de en yakın chunk aranır, top-1 skor kaydedilir. - Farklı threshold değerleri için confusion matrix hesaplanır: Pozitif soru + skor >= threshold -> TP (doğru yanıtlandı) Pozitif soru + skor < threshold -> FN (yanlışlıkla reddedildi) Negatif soru + skor < threshold -> TN (doğru reddedildi) Negatif soru + skor >= threshold -> FP (yanlışlıkla yanıtlandı / halüsinasyon riski) - En iyi accuracy/F1'i veren threshold önerilir. - Ayrıca pozitif sorularda, dönen chunk'ın url'i beklenen url ile eşleşiyor mu (retrieval doğruluğu) kontrol edilir. Çalıştırma: EMBEDDING_BACKEND=mock python eval/run_eval.py """ import sys import os import json sys.path.append(os.path.join(os.path.dirname(__file__), "..")) from src import config from src.vector_store import VectorStore from src.embedder import get_embedder THRESHOLD_SWEEP = [round(x * 0.05, 2) for x in range(1, 20)] # 0.05, 0.10, ..., 0.95 def load_questions(): path = os.path.join(os.path.dirname(__file__), "test_questions.json") with open(path, encoding="utf-8") as f: data = json.load(f) return data["questions"] def run_raw_search(questions, store, embedder, top_k=None): """Her soru için (threshold uygulamadan) top-1 skoru ve retrieval bilgisini toplar.""" results = [] for q in questions: query_vector = embedder.embed([q["question"]])[0] hits = store.query(query_vector, top_k=top_k or config.TOP_K) top = hits[0] if hits else {"score": 0.0, "url": None, "chunk_text": ""} results.append({ "id": q["id"], "type": q["type"], "question": q["question"], "expected_source_url": q.get("expected_source_url"), "top_score": top["score"], "top_url": top.get("url"), "top_chunk_preview": top.get("chunk_text", "")[:80], }) return results def confusion_matrix_at_threshold(results, threshold): tp = fn = tn = fp = 0 for r in results: answered = r["top_score"] >= threshold if r["type"] == "positive": if answered: tp += 1 else: fn += 1 else: # negative if answered: fp += 1 else: tn += 1 total = tp + fn + tn + fp accuracy = (tp + tn) / total if total else 0.0 precision = tp / (tp + fp) if (tp + fp) else 0.0 recall = tp / (tp + fn) if (tp + fn) else 0.0 f1 = 2 * precision * recall / (precision + recall) if (precision + recall) else 0.0 return {"threshold": threshold, "tp": tp, "fn": fn, "tn": tn, "fp": fp, "accuracy": accuracy, "precision": precision, "recall": recall, "f1": f1} def sweep_thresholds(results): return [confusion_matrix_at_threshold(results, t) for t in THRESHOLD_SWEEP] def retrieval_accuracy(results): """Pozitif sorularda, en yüksek skorlu chunk'ın url'i beklenen url ile eşleşiyor mu?""" positives = [r for r in results if r["type"] == "positive"] correct = sum(1 for r in positives if r["top_url"] == r["expected_source_url"]) return correct, len(positives) def main(): print(f"Backend: {config.EMBEDDING_BACKEND} | Chroma: {config.CHROMA_PERSIST_DIR}\n") questions = load_questions() store = VectorStore() embedder = get_embedder() print(f"Toplam {len(questions)} soru çalıştırılıyor " f"({sum(1 for q in questions if q['type']=='positive')} pozitif, " f"{sum(1 for q in questions if q['type']=='negative')} negatif)...\n") results = run_raw_search(questions, store, embedder) # --- Ham sonuçlar --- print(f"{'ID':5} {'Tip':9} {'Skor':7} {'Doğru URL mü?':14} Soru") print("-" * 100) for r in results: url_match = "" if r["type"] == "positive": url_match = "EVET" if r["top_url"] == r["expected_source_url"] else "HAYIR" print(f"{r['id']:5} {r['type']:9} {r['top_score']:.3f} {url_match:14} {r['question'][:60]}") correct, total_pos = retrieval_accuracy(results) print(f"\nRetrieval doğruluğu (pozitif sorularda doğru kaynağı bulma): {correct}/{total_pos}") # --- Threshold sweep --- print("\n" + "=" * 70) print("THRESHOLD SWEEP") print("=" * 70) print(f"{'Thr':6} {'TP':4} {'FN':4} {'TN':4} {'FP':4} {'Acc':6} {'Prec':6} {'Rec':6} {'F1':6}") sweep = sweep_thresholds(results) for s in sweep: print(f"{s['threshold']:.2f} {s['tp']:4} {s['fn']:4} {s['tn']:4} {s['fp']:4} " f"{s['accuracy']:.3f} {s['precision']:.3f} {s['recall']:.3f} {s['f1']:.3f}") best = max(sweep, key=lambda s: (s["f1"], s["accuracy"])) print(f"\n>>> Önerilen threshold (en iyi F1): {best['threshold']} " f"(accuracy={best['accuracy']:.3f}, precision={best['precision']:.3f}, recall={best['recall']:.3f})") write_report(results, sweep, best, correct, total_pos) def write_report(results, sweep, best, correct, total_pos): path = os.path.join(os.path.dirname(__file__), "eval_results.md") lines = [ "# Eşik (Threshold) Analizi Sonuçları", "", f"- Embedding backend: `{config.EMBEDDING_BACKEND}`", f"- Toplam soru: {len(results)} " f"({sum(1 for r in results if r['type']=='positive')} pozitif, " f"{sum(1 for r in results if r['type']=='negative')} negatif)", f"- Retrieval doğruluğu (pozitif sorularda doğru url): {correct}/{total_pos}", f"- **Önerilen threshold: {best['threshold']}** " f"(F1={best['f1']:.3f}, accuracy={best['accuracy']:.3f}, " f"precision={best['precision']:.3f}, recall={best['recall']:.3f})", "", "## Threshold Sweep Tablosu", "", "| Threshold | TP | FN | TN | FP | Accuracy | Precision | Recall | F1 |", "|---|---|---|---|---|---|---|---|---|", ] for s in sweep: lines.append( f"| {s['threshold']:.2f} | {s['tp']} | {s['fn']} | {s['tn']} | {s['fp']} | " f"{s['accuracy']:.3f} | {s['precision']:.3f} | {s['recall']:.3f} | {s['f1']:.3f} |" ) lines += ["", "## Soru Bazlı Ham Sonuçlar", "", "| ID | Tip | Top Score | Top URL Doğru mu | Soru |", "|---|---|---|---|---|"] for r in results: url_match = "-" if r["type"] == "positive": url_match = "EVET" if r["top_url"] == r["expected_source_url"] else "HAYIR" lines.append(f"| {r['id']} | {r['type']} | {r['top_score']:.3f} | {url_match} | {r['question']} |") with open(path, "w", encoding="utf-8") as f: f.write("\n".join(lines)) print(f"\nRapor yazıldı: {path}") if __name__ == "__main__": main()