api/rag.py

import os
import json
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.decomposition import TruncatedSVD
from sklearn.pipeline import Pipeline

DATA_PATH = os.path.join(os.path.dirname(__file__), '..', 'data', 'raw', 'train.json')
RAG_DIMS  = 512

_pipeline: Pipeline | None = None
_pgvector_ready = False


# ── helpers ──────────────────────────────────────────────────────────────────

def _database_url() -> str:
    return os.environ.get(
        'DATABASE_URL',
        'postgresql://mlflow:mlflow123@localhost:5432/mlflow_db'
    )


def _connect():
    import psycopg2
    from pgvector.psycopg2 import register_vector
    conn = psycopg2.connect(_database_url())
    register_vector(conn)
    return conn


def _fit_pipeline(texts: list[str]) -> Pipeline:
    pipe = Pipeline([
        ('tfidf', TfidfVectorizer(max_features=5000, stop_words='english', ngram_range=(1, 2))),
        ('svd',   TruncatedSVD(n_components=RAG_DIMS, random_state=42)),
    ])
    pipe.fit(texts)
    return pipe


def _embed(pipe: Pipeline, texts: list[str]) -> np.ndarray:
    vecs = pipe.transform(texts)
    norms = np.linalg.norm(vecs, axis=1, keepdims=True)
    norms[norms == 0] = 1.0
    return (vecs / norms).astype(np.float32)


# ── startup ───────────────────────────────────────────────────────────────────

def load_rag_data(_main_vectorizer=None) -> None:
    global _pipeline, _pgvector_ready

    if not os.path.exists(DATA_PATH):
        print(f"RAG: {DATA_PATH} not found — retrieval disabled.")
        return

    try:
        with open(DATA_PATH, 'r', encoding='utf-8') as f:
            data = json.load(f)
    except Exception as e:
        print(f"RAG: failed to load data — {e}")
        return

    statements = [str(item.get('statement', '')) for item in data]
    labels     = [str(item.get('label',     '')) for item in data]
    reasons    = [str(item.get('reason',    '')) for item in data]

    # Always (re)fit the pipeline — fast enough on 10K rows
    print("RAG: fitting LSA pipeline (TF-IDF 5 000 → SVD 512)…")
    _pipeline = _fit_pipeline(statements)

    db_url = _database_url()
    print(f"RAG: connecting to {db_url.split('@')[-1]}…")  # hide credentials

    try:
        conn = _connect()
        cur  = conn.cursor()

        cur.execute("SELECT COUNT(*) FROM rag_claims;")
        count = cur.fetchone()[0]

        if count == 0:
            print(f"RAG: inserting {len(statements)} claims into pgvector…")
            vectors = _embed(_pipeline, statements)

            from psycopg2.extras import execute_values
            rows = [
                (statements[i], labels[i], reasons[i][:500], vectors[i])
                for i in range(len(statements))
            ]
            execute_values(
                cur,
                "INSERT INTO rag_claims (statement, label, reason, embedding) VALUES %s",
                rows,
                template="(%s, %s, %s, %s)",
            )
            conn.commit()
            print(f"RAG: {len(rows)} claims indexed in pgvector ✓")
        else:
            print(f"RAG: {count} claims already in pgvector — skipping insert.")

        cur.close()
        conn.close()
        _pgvector_ready = True
        print("RAG: ready ✓")

    except Exception as e:
        print(f"RAG ERROR: {e}")
        print("RAG: retrieval disabled — run 'docker compose down -v && docker compose up --build' to reset.")
        _pgvector_ready = False


# ── retrieval ─────────────────────────────────────────────────────────────────

def retrieve_similar(claim: str, predicted_label: str, top_k: int = 3) -> list:
    if not _pgvector_ready or _pipeline is None:
        return []

    try:
        query_vec = _embed(_pipeline, [claim])[0]

        conn = _connect()
        cur  = conn.cursor()
        cur.execute(
            """
            SELECT statement, label, reason,
                   1 - (embedding <=> %s) AS similarity
            FROM   rag_claims
            WHERE  label = %s
            ORDER  BY embedding <=> %s
            LIMIT  %s
            """,
            (query_vec, predicted_label, query_vec, top_k),
        )
        rows = cur.fetchall()
        cur.close()
        conn.close()

        return [
            {
                'statement':  row[0],
                'label':      row[1],
                'reason':     row[2],
                'similarity': float(row[3]),
            }
            for row in rows
        ]
    except Exception as e:
        print(f"RAG: retrieval error — {e}")
        return []


# ── generation (NVIDIA NIM) ───────────────────────────────────────────────────

_NIM_BASE_URL = "https://integrate.api.nvidia.com/v1"
_NIM_MODEL    = "meta/llama-3.1-8b-instruct"


def _generate_justification(claim: str, predicted_label: str, similar: list) -> str | None:
    if not similar:
        return None

    nim_key = os.environ.get('NVIDIA_API_KEY')
    if not nim_key:
        print("RAG: NVIDIA_API_KEY not set — justification disabled.")
        return None

    try:
        from openai import OpenAI
        client = OpenAI(base_url=_NIM_BASE_URL, api_key=nim_key)
    except ImportError:
        return None

    label_fr = {
        'true':        'VRAI',
        'mostly-true': 'MAJORITAIREMENT VRAI',
        'half-true':   'PARTIELLEMENT VRAI',
        'barely-true': 'À PEINE VRAI',
        'false':       'FAUX',
        'pants-fire':  'TOTALEMENT FAUX',
    }.get(predicted_label, predicted_label.upper())

    examples = []
    for i, s in enumerate(similar[:3], 1):
        examples.append(
            f"Exemple {i} (similarité={s['similarity']:.2f}) :\n"
            f"  Déclaration : {s['statement']}\n"
            f"  Raison : {(s['reason'] or '')[:250]}"
        )

    prompt = (
        f"Tu es un assistant de vérification des faits.\n"
        f"Déclaration : « {claim} »\n"
        f"Classification : {label_fr} ({predicted_label})\n\n"
        f"Exemples similaires (même classification) dans la base LIAR :\n\n"
        + '\n\n'.join(examples) +
        f"\n\nEn 2-3 phrases concises en français, justifie pourquoi cette déclaration est"
        f" classée « {predicted_label} » en t'appuyant sur les similitudes avec ces exemples."
        " Réponds directement sans titre ni introduction."
    )

    try:
        resp = client.chat.completions.create(
            model=_NIM_MODEL,
            messages=[{"role": "user", "content": prompt}],
            max_tokens=350,
            temperature=0.3,
        )
        return resp.choices[0].message.content.strip()
    except Exception as exc:
        print(f"RAG NIM generation error: {exc}")
        return None


# ── public API ────────────────────────────────────────────────────────────────

def get_rag_result(claim: str, predicted_label: str, _vectorizer=None) -> tuple[list, str | None]:
    similar       = retrieve_similar(claim, predicted_label)
    justification = _generate_justification(claim, predicted_label, similar)

    evidence = [
        {
            'statement':  s['statement'],
            'label':      s['label'],
            'similarity': round(s['similarity'], 3),
            'reason':     (s['reason'] or '')[:300] or None,
        }
        for s in similar
    ]
    return evidence, justification