ablation_study_squad.py

# ==============================================================================
#  AetherMap — ABLATION STUDY v2 (SQuAD EN)
#  Usa SQuAD v1.1 original (inglês) com Q&A humanas
#  Para rodar no Google Colab contra a API do HF Space
# ==============================================================================

# %%
# !pip install requests pandas numpy matplotlib datasets -q

# %%
import os
import json
import time
import random
import requests
import pandas as pd
import numpy as np
import tempfile
from typing import List, Dict, Any
from datasets import load_dataset

# ==============================================================================
# CONFIGURAÇÃO
# ==============================================================================
OPENROUTER_API_KEY = "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"  # 👈 SUA KEY OPENROUTER
AETHERMAP_URL = "https://madras1-aethermap.hf.space"

# Modelo pro LLM-as-Judge (avaliador)
EVALUATOR_MODEL = "nvidia/nemotron-3-nano-30b-a3b:free"

DELAY_BETWEEN_LLM_CALLS = 2.0  # Segundos entre chamadas LLM

OPENROUTER_HEADERS = {
    "Authorization": f"Bearer {OPENROUTER_API_KEY}",
    "Content-Type": "application/json",
}

# Modos de ablation e seus componentes
ABLATION_MODES = {
    "faiss_only":     {"label": "FAISS Only",          "faiss": True,  "bm25": False, "rrf": False, "reranker": False, "expansion": False},
    "bm25_only":      {"label": "BM25 Only",           "faiss": False, "bm25": True,  "rrf": False, "reranker": False, "expansion": False},
    "hybrid":         {"label": "Hybrid (RRF)",        "faiss": True,  "bm25": True,  "rrf": True,  "reranker": False, "expansion": False},
    "hybrid_rerank":  {"label": "+ Reranker",          "faiss": True,  "bm25": True,  "rrf": True,  "reranker": True,  "expansion": False},
    "full":           {"label": "Full Pipeline",       "faiss": True,  "bm25": True,  "rrf": True,  "reranker": True,  "expansion": True},
}

print("✅ Configuração carregada")
print(f"🌐 API: {AETHERMAP_URL}")
print(f"🧑‍⚖️ Avaliador: {EVALUATOR_MODEL}")
print("📚 Dataset: SQuAD v1.1 (English)")

# %%
# ==============================================================================
# CARREGAR SQUAD PT-BR
# ==============================================================================

def carregar_squad_pt(n_contexts: int = 300, n_queries: int = 30, seed: int = 42) -> tuple:
    """
    Carrega SQuAD v1.1 PT-BR e extrai contextos + Q&A pairs.
    
    Returns:
        contexts: Lista de parágrafos únicos para indexar
        qa_pairs: Lista de {pergunta, resposta, contexto_id} para testar
    """
    print("📥 Baixando SQuAD v1.1 (inglês) do HuggingFace...")
    ds = load_dataset("rajpurkar/squad", split="validation")  # validation = 10k exemplos, mais rápido
    print(f"   📊 Total: {len(ds)} exemplos no dataset")
    print(f"   📋 Colunas: {ds.column_names}")
    
    # SQuAD original: id, title, context, question, answers
    ctx_col = "context"
    q_col = "question"
    ans_col = "answers"
    title_col = "title"
    
    print(f"   🔍 Usando: context='{ctx_col}', question='{q_col}', answers='{ans_col}', title='{title_col}'")
    
    # Extrair contextos únicos
    context_set = {}
    for item in ds:
        ctx = str(item[ctx_col]).strip()
        if ctx and len(ctx) > 50 and ctx not in context_set:
            context_set[ctx] = {
                "text": ctx,
                "title": str(item.get(title_col, "")) if title_col else "",
            }
    
    all_contexts = list(context_set.values())
    print(f"   📄 {len(all_contexts)} contextos únicos encontrados")
    
    # Amostrar contextos
    random.seed(seed)
    sampled_contexts = random.sample(all_contexts, min(n_contexts, len(all_contexts)))
    sampled_texts = set(c["text"] for c in sampled_contexts)
    
    print(f"   ✂️ {len(sampled_contexts)} contextos amostrados")
    
    # Coletar Q&A pairs que pertencem aos contextos amostrados
    qa_candidates = []
    for item in ds:
        ctx = str(item[ctx_col]).strip()
        if ctx in sampled_texts:
            # Extrair resposta (formato pode ser dict com "text" ou string direta)
            if ans_col:
                ans_raw = item[ans_col]
                if isinstance(ans_raw, dict) and "text" in ans_raw:
                    answer_texts = ans_raw["text"]
                    if isinstance(answer_texts, list) and answer_texts:
                        answer = answer_texts[0].strip()
                    else:
                        answer = str(answer_texts).strip()
                elif isinstance(ans_raw, list) and ans_raw:
                    answer = str(ans_raw[0]).strip()
                else:
                    answer = str(ans_raw).strip()
            else:
                answer = ""
            
            question = str(item[q_col]).strip()
            
            if question and answer and len(answer) > 2:
                qa_candidates.append({
                    "pergunta": question,
                    "resposta": answer,
                    "contexto": ctx[:200],
                })
    
    print(f"   ❓ {len(qa_candidates)} Q&A pairs nos contextos amostrados")
    
    # Amostrar queries diversificadas por contexto
    random.shuffle(qa_candidates)
    seen_contexts = set()
    diverse_queries = []
    for qa in qa_candidates:
        ctx_key = qa["contexto"][:100]
        if ctx_key not in seen_contexts:
            diverse_queries.append(qa)
            seen_contexts.add(ctx_key)
        if len(diverse_queries) >= n_queries:
            break
    
    # Completar se necessário
    if len(diverse_queries) < n_queries:
        remaining = [q for q in qa_candidates if q not in diverse_queries]
        diverse_queries.extend(remaining[:n_queries - len(diverse_queries)])
    
    print(f"   ✅ {len(diverse_queries)} queries selecionadas (diversificadas por contexto)")
    
    return sampled_contexts, diverse_queries


# %%
# ==============================================================================
# FUNÇÕES AUXILIARES
# ==============================================================================

def call_llm(prompt: str, model: str, max_tokens: int = 1000) -> str:
    """Chama LLM via OpenRouter usando requests direto."""
    time.sleep(DELAY_BETWEEN_LLM_CALLS)
    try:
        payload = {
            "model": model,
            "messages": [{"role": "user", "content": prompt}],
            "max_tokens": max_tokens,
            "temperature": 0.3,
            "reasoning": {
                "exclude": True  # Reasoning vai pra campo separado, content fica limpo
            },
        }

        resp = requests.post(
            "https://openrouter.ai/api/v1/chat/completions",
            headers=OPENROUTER_HEADERS,
            json=payload,
            timeout=90
        )

        if resp.status_code != 200:
            print(f"   ⚠️ LLM ({model}): HTTP {resp.status_code}")
            print(f"   📝 {resp.text[:300]}")
            return ""

        data = resp.json()

        choices = data.get("choices", [])
        if not choices:
            print(f"   ⚠️ LLM ({model}): Sem choices")
            print(f"   📝 Raw: {json.dumps(data)[:400]}")
            return ""

        message = choices[0].get("message", {})
        content = message.get("content") or ""

        if not content:
            print(f"   ⚠️ LLM ({model}): content vazio")
            print(f"   📝 Message keys: {list(message.keys())}")
            print(f"   📝 Raw: {json.dumps(message)[:400]}")
            return ""

        return content.strip()

    except Exception as e:
        print(f"   ⚠️ Erro LLM ({model}): {type(e).__name__}: {e}")
        time.sleep(5)
        return ""


def avaliar_resposta(pergunta: str, resposta_rag: str, resposta_esperada: str) -> Dict[str, float]:
    """
    Avalia qualidade da resposta do RAG comparando com ground truth do SQuAD.
    Retorna scores de correção, completude e relevância (0-1).
    """
    prompt = f"""Avalie a resposta do sistema comparando com a resposta de referência (ground truth humana).

PERGUNTA: {pergunta}
RESPOSTA DE REFERÊNCIA (humana): {resposta_esperada[:300]}
RESPOSTA DO SISTEMA (RAG): {resposta_rag[:400]}

Dê notas de 0.0 a 1.0 para:
- correcao: A informação está factualmente correta comparada à referência?
- completude: A resposta cobre todos os pontos da referência?
- relevancia: A resposta é diretamente relevante à pergunta?

Retorne APENAS JSON: {{"correcao": X.X, "completude": X.X, "relevancia": X.X}}"""

    response = call_llm(prompt, EVALUATOR_MODEL, max_tokens=500)

    try:
        if "{" in response:
            json_start = response.find("{")
            json_end = response.rfind("}")
            if json_end > json_start:
                result = json.loads(response[json_start:json_end + 1])
                return {
                    "correcao": float(result.get("correcao", 0.5)),
                    "completude": float(result.get("completude", 0.5)),
                    "relevancia": float(result.get("relevancia", 0.5)),
                }
    except Exception as e:
        print(f"   ⚠️ Erro parsing avaliação: {e}")
        print(f"   📝 Response: {response[:200]}")
    return {"correcao": 0.5, "completude": 0.5, "relevancia": 0.5}


# %%
# ==============================================================================
# UPLOAD DO DATASET
# ==============================================================================

def upload_contexts(contexts: List[Dict], text_column: str = "texto") -> str:
    """Faz upload dos contextos SQuAD para o AetherMap."""
    
    # Criar CSV com os contextos
    df = pd.DataFrame({
        "texto": [c["text"] for c in contexts],
        "titulo": [c["title"] for c in contexts],
    })
    
    with tempfile.NamedTemporaryFile(mode='w', suffix='.csv', delete=False, encoding='utf-8') as f:
        df.to_csv(f, index=False)
        temp_path = f.name
    
    print(f"📤 Upload de {len(contexts)} contextos SQuAD...")
    
    with open(temp_path, "rb") as f:
        response = requests.post(
            f"{AETHERMAP_URL}/process/",
            files={"file": ("squad_contexts.csv", f)},
            data={
                "n_samples": len(contexts),
                "text_column": text_column,
                "fast_mode": "true"
            },
            timeout=300
        )
    os.unlink(temp_path)
    
    if response.status_code != 200:
        raise Exception(f"Erro no upload: {response.text[:200]}")
    
    result = response.json()
    job_id = result["job_id"]
    meta = result["metadata"]
    
    print(f"✅ Job criado: {job_id[:8]}...")
    print(f"   📊 {meta['num_documents_processed']} docs | {meta['num_clusters_found']} clusters")
    
    return job_id


# %%
# ==============================================================================
# BUSCA COM MODO DE ABLATION
# ==============================================================================

def search_with_mode(job_id: str, query: str, mode: str) -> Dict:
    """Executa busca com um modo de ablation específico."""
    start = time.time()

    try:
        response = requests.post(
            f"{AETHERMAP_URL}/search/",
            data={
                "query": query,
                "job_id": job_id,
                "ablation_mode": mode
            },
            timeout=300  # 5 min — Full Pipeline com 300 docs pode demorar
        )

        latency = time.time() - start

        if response.status_code != 200:
            return {"error": response.text[:100], "latency": latency}

        result = response.json()
        result["latency"] = latency
        return result
    
    except requests.exceptions.Timeout:
        latency = time.time() - start
        print(f"   ⏰ Timeout após {latency:.0f}s")
        return {"error": f"Timeout ({latency:.0f}s)", "latency": latency}
    except Exception as e:
        latency = time.time() - start
        print(f"   ⚠️ Erro: {type(e).__name__}: {e}")
        return {"error": str(e)[:100], "latency": latency}


# %%
# ==============================================================================
# ABLATION STUDY PRINCIPAL
# ==============================================================================

def run_ablation_study(
    n_contexts: int = 300,
    n_queries: int = 30
):
    """
    Executa ablation study completo com SQuAD PT-BR.
    
    Args:
        n_contexts: Número de contextos para indexar
        n_queries: Número de queries para testar
    """
    print("=" * 60)
    print("🧪 AETHERMAP ABLATION STUDY v2 — SQuAD EN")
    print("=" * 60)
    
    # 1. Carregar SQuAD
    print("\n📚 FASE 1: Carregando SQuAD PT-BR")
    contexts, queries = carregar_squad_pt(n_contexts, n_queries)
    
    # 2. Upload
    print(f"\n📦 FASE 2: Upload dos contextos")
    job_id = upload_contexts(contexts)
    
    # 3. Mostrar queries
    print(f"\n❓ FASE 3: {len(queries)} queries do SQuAD EN (ground truth humana)")
    for i, q in enumerate(queries):
        print(f"   [{i+1}] Q: {q['pergunta'][:70]}...")
        print(f"        A: {q['resposta'][:70]}...")
    
    # 4. Rodar cada modo
    print(f"\n🧪 FASE 4: Testando {len(ABLATION_MODES)} modos de ablation...")
    
    all_results = {}
    
    for mode_key, mode_info in ABLATION_MODES.items():
        mode_label = mode_info["label"]
        print(f"\n{'─' * 40}")
        print(f"🔬 Modo: {mode_label} ({mode_key})")
        print(f"   Componentes: ", end="")
        components = []
        if mode_info["faiss"]: components.append("FAISS")
        if mode_info["bm25"]: components.append("BM25")
        if mode_info["rrf"]: components.append("RRF")
        if mode_info["reranker"]: components.append("Reranker")
        if mode_info["expansion"]: components.append("QueryExpansion")
        print(" + ".join(components))
        
        scores = []
        latencies = []
        
        for i, q in enumerate(queries):
            pergunta = q["pergunta"]
            esperada = q["resposta"]
            
            # Buscar
            result = search_with_mode(job_id, pergunta, mode_key)
            
            if "error" in result:
                print(f"   [{i+1}] ❌ Erro: {result['error'][:50]}")
                scores.append({"correcao": 0, "completude": 0, "relevancia": 0})
                latencies.append(result.get("latency", 0))
                continue
            
            resposta = result.get("summary", "")
            latency = result.get("latency", 0)
            latencies.append(latency)
            
            # Avaliar
            aval = avaliar_resposta(pergunta, resposta, esperada)
            scores.append(aval)
            
            avg = (aval["correcao"] + aval["completude"] + aval["relevancia"]) / 3
            emoji = "✓" if avg >= 0.6 else "○"
            print(f"   [{i+1}] {emoji} C:{aval['correcao']:.2f} Com:{aval['completude']:.2f} R:{aval['relevancia']:.2f} | {latency:.1f}s")
            print(f"         Q: {pergunta[:80]}")
            print(f"         RAG: {resposta[:150]}...")
            if avg < 0.3:
                print(f"         ⚠️ Esperado: {esperada[:100]}")
            
            time.sleep(0.5)  # Rate limiting suave
        
        # Calcular médias
        avg_correcao = np.mean([s["correcao"] for s in scores])
        avg_completude = np.mean([s["completude"] for s in scores])
        avg_relevancia = np.mean([s["relevancia"] for s in scores])
        avg_score = (avg_correcao + avg_completude + avg_relevancia) / 3
        avg_latency = np.mean(latencies)
        
        all_results[mode_key] = {
            "label": mode_label,
            "avg_score": avg_score,
            "correcao": avg_correcao,
            "completude": avg_completude,
            "relevancia": avg_relevancia,
            "avg_latency": avg_latency,
            "n_queries": len(queries),
            "components": components,
        }
        
        print(f"   📊 Score médio: {avg_score:.3f} | Latência: {avg_latency:.1f}s")
    
    # 5. Resultado Final
    print_results(all_results)
    plot_results(all_results)
    
    return all_results


# %%
# ==============================================================================
# VISUALIZAÇÃO DOS RESULTADOS
# ==============================================================================

def print_results(results: Dict):
    """Imprime tabela comparativa dos resultados."""
    print("\n" + "=" * 70)
    print("📊 RESULTADO DO ABLATION STUDY — SQuAD EN")
    print("=" * 70)

    print(f"\n{'Modo':<20} {'Score':>7} {'Correção':>9} {'Complet.':>9} {'Relev.':>7} {'Latência':>9} {'Δ Score':>8}")
    print("─" * 70)

    baseline_score = results.get("faiss_only", {}).get("avg_score", 0)
    best_score = max(r["avg_score"] for r in results.values())

    for mode_key in ABLATION_MODES.keys():
        if mode_key not in results:
            continue
        r = results[mode_key]
        delta = r["avg_score"] - baseline_score
        delta_str = f"+{delta:.3f}" if delta > 0 else f"{delta:.3f}"
        trophy = " 🏆" if r["avg_score"] == best_score else ""

        print(f"{r['label']:<20} {r['avg_score']:>7.3f} {r['correcao']:>9.3f} {r['completude']:>9.3f} {r['relevancia']:>7.3f} {r['avg_latency']:>8.1f}s {delta_str:>8}{trophy}")

    # Insights
    print(f"\n{'─' * 70}")
    print("💡 INSIGHTS:")

    if "faiss_only" in results and "hybrid" in results:
        delta = results["hybrid"]["avg_score"] - results["faiss_only"]["avg_score"]
        base = max(results['faiss_only']['avg_score'], 0.001)
        print(f"   Hybrid Search (RRF): {'+' if delta >= 0 else ''}{delta:.3f} vs FAISS only ({delta/base*100:+.1f}%)")

    if "hybrid" in results and "hybrid_rerank" in results:
        delta = results["hybrid_rerank"]["avg_score"] - results["hybrid"]["avg_score"]
        base = max(results['hybrid']['avg_score'], 0.001)
        print(f"   + Reranker:          {'+' if delta >= 0 else ''}{delta:.3f} vs Hybrid ({delta/base*100:+.1f}%)")

    if "hybrid_rerank" in results and "full" in results:
        delta = results["full"]["avg_score"] - results["hybrid_rerank"]["avg_score"]
        base = max(results['hybrid_rerank']['avg_score'], 0.001)
        print(f"   + Query Expansion:   {'+' if delta >= 0 else ''}{delta:.3f} vs Hybrid+Rerank ({delta/base*100:+.1f}%)")

    # Comparar n_queries
    n = list(results.values())[0]["n_queries"]
    print(f"\n   📏 N = {n} queries | Dataset: SQuAD v1.1 EN (ground truth humana)")


def plot_results(results: Dict):
    """Gera gráfico de barras com os resultados."""
    try:
        import matplotlib.pyplot as plt
        import matplotlib
        matplotlib.rcParams['figure.facecolor'] = '#0d1117'
        matplotlib.rcParams['axes.facecolor'] = '#161b22'
        matplotlib.rcParams['text.color'] = '#c9d1d9'
        matplotlib.rcParams['axes.labelcolor'] = '#c9d1d9'
        matplotlib.rcParams['xtick.color'] = '#8b949e'
        matplotlib.rcParams['ytick.color'] = '#8b949e'
    except ImportError:
        print("⚠️ matplotlib não disponível. Pulando gráfico.")
        return

    labels = [results[k]["label"] for k in ABLATION_MODES if k in results]
    scores = [results[k]["avg_score"] for k in ABLATION_MODES if k in results]
    latencies = [results[k]["avg_latency"] for k in ABLATION_MODES if k in results]

    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 6))

    # Cores neon progressivas
    colors = ['#6366f1', '#818cf8', '#34d399', '#fbbf24', '#f472b6'][:len(labels)]

    # Gráfico 1: Scores
    bars1 = ax1.bar(labels, scores, color=colors, edgecolor='#30363d', linewidth=1.5)
    ax1.set_title('Quality Score por Modo', fontsize=14, fontweight='bold', pad=15)
    ax1.set_ylabel('Score Médio (0-1)')
    ax1.set_ylim(0, 1.05)
    ax1.grid(axis='y', alpha=0.15, color='#8b949e')
    for bar, score in zip(bars1, scores):
        ax1.text(bar.get_x() + bar.get_width()/2., bar.get_height() + 0.02,
                f'{score:.3f}', ha='center', va='bottom', fontweight='bold', fontsize=11)

    # Gráfico 2: Latência
    bars2 = ax2.bar(labels, latencies, color=colors, edgecolor='#30363d', linewidth=1.5)
    ax2.set_title('Latência por Modo', fontsize=14, fontweight='bold', pad=15)
    ax2.set_ylabel('Latência Média (s)')
    ax2.grid(axis='y', alpha=0.15, color='#8b949e')
    for bar, lat in zip(bars2, latencies):
        ax2.text(bar.get_x() + bar.get_width()/2., bar.get_height() + 0.1,
                f'{lat:.1f}s', ha='center', va='bottom', fontweight='bold', fontsize=11)

    plt.xticks(rotation=25, ha='right')
    plt.suptitle('AetherMap RAG — Ablation Study (SQuAD EN)', fontsize=16, fontweight='bold', y=1.02, color='#f0f6fc')
    plt.tight_layout()
    plt.savefig('ablation_results_squad.png', dpi=150, bbox_inches='tight',
                facecolor='#0d1117', edgecolor='none')
    plt.show()
    print("📈 Gráfico salvo em ablation_results_squad_en.png")


# %%
# ==============================================================================
# EXECUTAR
# ==============================================================================

resultados = run_ablation_study(
    n_contexts=300,
    n_queries=30
)