File size: 6,008 Bytes
39acd95 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 | """
AIBalance API β recommandation sobre par usage
Version legere : charge des fichiers pre-calcules (pas de parquet 4 Go).
Pre-requis : lancer export_for_api.py une fois en local.
Endpoint : POST /recommend {prompt, mode, top_n}
"""
from flask import Flask, request, jsonify
from flask_cors import CORS
from sentence_transformers import SentenceTransformer
import pandas as pd, re, time, os, pickle, json, joblib
app = Flask(__name__)
CORS(app)
# ββ Config ββ
EMBEDDING_MODEL = "Lajavaness/bilingual-embedding-small"
DELTA = 0.03
MEDIAN_ENERGY = 0.66
MODEL_PATH = "outputs_v3/api_clf.joblib"
# ββ Globals ββ
embedder = None
clf = None
reco_table = {}
profil_map = {}
FAMILY_MAP = {
'Natural Science & Formal Science & Technology': 'code',
'Education': 'education', 'Arts & Culture': 'creation', 'Arts': 'creation',
'Entertainment & Travel & Hobby': 'creation', 'Food & Drink & Cooking': 'creation',
}
def get_task_family(cat):
return FAMILY_MAP.get(cat, 'analysis' if 'Business' in cat or 'Law' in cat or 'Politic' in cat else 'education' if 'Education' in cat else 'default')
def extract_complexity(prompt):
p = str(prompt)
return {
'n_words': len(p.split()),
'is_code': int(bool(re.search(r'\b(code|python|java|sql|html|css|script|function|class|def |import)\b', p, re.I))),
'n_questions': p.count('?'),
}
def boot():
"""Charge les fichiers pre-calcules + embedder + classifieur."""
global embedder, clf, reco_table, profil_map
print("Boot API (mode leger)...")
t0 = time.time()
# embedder
device = 'cpu'
print(f" Device : {device}")
embedder = SentenceTransformer(EMBEDDING_MODEL, trust_remote_code=True, device=device)
# classifieur
clf_obj = joblib.load(MODEL_PATH)
print(f" Classifieur charge depuis {MODEL_PATH}")
# reco_table pre-calculee
with open('outputs_v3/api_reco_table.pkl', 'rb') as f:
reco_table.update(pickle.load(f))
print(f" Reco table : {len(reco_table)} categories")
# profil map
with open('outputs_v3/api_profil_map.json', 'r') as f:
profil_map.update(json.load(f))
print(f"Boot OK ({time.time() - t0:.1f}s)")
return clf_obj
def classify_prompt(prompt):
emb = embedder.encode([prompt], convert_to_numpy=True, normalize_embeddings=True)
return clf.predict(emb)[0]
def recommander(prompt, top_n=3, mode='balanced'):
MODES = {'performance': (0.8, 0.2), 'green': (0.2, 0.8), 'balanced': (0.5, 0.5)}
alpha, beta = MODES.get(mode, (0.5, 0.5))
category = classify_prompt(prompt)
family = get_task_family(category)
cx = extract_complexity(prompt)
if category in reco_table:
candidates = reco_table[category].copy()
qual_col = 'combined_quality'
else:
found = False
for orig in reco_table:
if get_task_family(orig) == get_task_family(category):
candidates = reco_table[orig].copy()
qual_col = 'combined_quality'
found = True
break
if not found:
candidates = pd.DataFrame()
qual_col = 'win_rate'
if len(candidates) == 0:
return {'error': 'Pas de candidats pour cette categorie', 'category': category}
# score mode
q = candidates[qual_col]
q_norm = (q - q.min()) / (q.max() - q.min()) if q.max() > q.min() else 0.5
e = candidates['wh_per_1k_tok']
s_norm = 1 - (e - e.min()) / (e.max() - e.min()) if e.max() > e.min() else 0.5
candidates['mode_score'] = alpha * q_norm + beta * s_norm
# filtre DELTA
best_q = candidates[qual_col].max()
equivalent = candidates[candidates[qual_col] >= best_q - DELTA]
# filtre complexite
is_complex = cx['n_words'] > 100 or cx['is_code'] or cx['n_questions'] > 2
if is_complex:
wr_col = 'win_rate_cat' if 'win_rate_cat' in equivalent.columns else 'win_rate'
median_wr = candidates[wr_col].median()
complex_ok = equivalent[equivalent[wr_col] >= median_wr]
if len(complex_ok) > 0:
equivalent = complex_ok
top = equivalent.nlargest(top_n, 'mode_score')
# resultats
results = []
for _, r in top.iterrows():
wr = r.get('win_rate_cat', r.get('win_rate', 0))
cq = r.get(qual_col, wr)
energy = r['wh_per_1k_tok']
gain_vs_median = (1 - energy / MEDIAN_ENERGY) * 100 if MEDIAN_ENERGY > energy else 0
results.append({
'model': r['model'],
'energy_wh': round(energy, 3),
'win_rate': round(float(wr) * 100, 1),
'quality': round(float(cq), 3),
'profil': profil_map.get(r['model'], '?'),
'gain_vs_median': round(gain_vs_median, 0),
})
return {
'category': category,
'family': family,
'complexity': 'complexe' if is_complex else 'simple',
'n_words': cx['n_words'],
'mode': mode,
'alpha': alpha,
'beta': beta,
'n_candidates': len(candidates),
'n_equivalent': len(equivalent),
'delta': DELTA,
'recommendations': results,
}
@app.route('/recommend', methods=['POST'])
def api_recommend():
data = request.get_json()
prompt = data.get('prompt', '')
mode = data.get('mode', 'balanced')
top_n = data.get('top_n', 3)
if not prompt.strip():
return jsonify({'error': 'Prompt vide'}), 400
result = recommander(prompt, top_n=top_n, mode=mode)
return jsonify(result)
@app.route('/health', methods=['GET'])
def health():
return jsonify({
'status': 'ok',
'categories': len(reco_table),
'models': sum(len(t) for t in reco_table.values()),
})
if __name__ == '__main__':
clf = boot()
port = int(os.environ.get('PORT', 7860))
print(f"\nAPI prete : http://localhost:{port}")
print(" POST /recommend {prompt, mode, top_n}")
print(" GET /health")
app.run(host='0.0.0.0', port=port, debug=False)
|