API_RapidAPI_test / services /analysis.py
Moncef Souri
fix(analysis): enforce 3-key JSON output and fix French placeholder responses
8cff4cb
Raw History Blame Contribute Delete
13.6 kB
# services/analysis.py
"""
Groq LLM analysis with:
β€’ singleton Groq client (created once, reused across all requests)
β€’ deduplication of detections before sending to Groq (fewer tokens β†’ faster)
β€’ in-memory response cache (identical detections β†’ instant reply)
β€’ hard timeout so slow API calls never block the endpoint
β€’ strict 3-key output: conseils, interpretation, routine
"""
import os
import re
import json
import time
import hashlib
import threading
from collections import defaultdict
from typing import Optional
from groq import Groq
from config.settings import GROQ_TIMEOUT, GROQ_MODEL, MAX_CACHE_SIZE
# ── Singleton Groq client ──────────────────────────────────────────────────────
# Created once on first use and reused across all requests.
# Avoids the overhead of building a new HTTP client per request.
_groq_client: Optional[Groq] = None
_groq_client_lock = threading.Lock()
def _get_client() -> Optional[Groq]:
global _groq_client
if _groq_client is not None:
return _groq_client
with _groq_client_lock:
# Double-check inside the lock (another thread may have initialized it)
if _groq_client is None:
api_key = os.environ.get("GROQ_API_KEY")
if not api_key:
return None
_groq_client = Groq(api_key=api_key)
print("[analysis] Groq client initialized βœ“")
return _groq_client
# ── Simple in-memory cache ─────────────────────────────────────────────────────
_cache: dict[str, dict] = {}
_cache_lock = threading.Lock()
def _cache_key(detections: list, language: str) -> str:
issues = sorted(d['class_name'] for d in detections)
return hashlib.md5(f"{','.join(issues)}:{language}".encode()).hexdigest()
def _store(key: str, value: dict) -> None:
with _cache_lock:
if len(_cache) >= MAX_CACHE_SIZE:
oldest = next(iter(_cache))
del _cache[oldest]
_cache[key] = value
# ── Deduplication ──────────────────────────────────────────────────────────────
def _deduplicate(detections: list) -> list:
"""
Collapse multiple detections of the same class into one entry,
keeping only the highest confidence score per class.
Example: 8x 'Acne Scar' detections β†’ 1x 'Acne Scar' (highest confidence)
This significantly reduces Groq prompt size and speeds up the API call.
"""
best: dict[str, float] = {}
for d in detections:
name = d['class_name']
if name not in best or d['confidence'] > best[name]:
best[name] = d['confidence']
# Return sorted by confidence descending so most prominent issues come first
return [
{'class_name': name, 'confidence': conf}
for name, conf in sorted(best.items(), key=lambda x: -x[1])
]
# ── Default fallback ───────────────────────────────────────────────────────────
def get_default_analysis(language: str = 'en', timeout: bool = False) -> dict:
if language == 'fr':
if timeout:
return {
"conseils": "L'analyse prend plus de temps que prΓ©vu. Veuillez rΓ©essayer.",
"interpretation": "Le système d'analyse IA est temporairement occupé.",
"routine": "Routine standard: nettoyage doux, hydratation et protection solaire quotidienne. Consultez la section des produits recommandΓ©s ci-dessous pour voir les solutions adaptΓ©es Γ  vos besoins spΓ©cifiques.",
}
return {
"conseils": "Consultez un dermatologue pour des conseils personnalisΓ©s.",
"interpretation": "Une analyse professionnelle est recommandée pour ces problèmes de peau.",
"routine": "Nettoyage doux matin et soir, hydratation quotidienne et protection solaire. Consultez la section des produits recommandΓ©s ci-dessous pour voir les solutions adaptΓ©es Γ  vos besoins spΓ©cifiques.",
}
if timeout:
return {
"conseils": "Analysis is taking longer than expected. Please try again.",
"interpretation": "The AI analysis system is temporarily busy.",
"routine": "Standard routine: gentle cleansing, moisturising, and daily sun protection. Check the recommended products section below to see solutions tailored to your specific needs.",
}
return {
"conseils": "Consult a dermatologist for personalised advice.",
"interpretation": "Professional analysis is recommended for these skin problems.",
"routine": "Follow a basic routine: cleansing, moisturising, and daily sun protection. Check the recommended products section below to see solutions tailored to your specific needs.",
}
# ── Placeholder guard ──────────────────────────────────────────────────────────
_PLACEHOLDERS = {
"conseils courts et pratiques", "brief practical advice",
"causes et gravité brèves", "brief causes and severity",
"routine quotidienne courte", "brief daily routine",
"[tes conseils ici]", "[ton interpretation ici]", "[ta routine ici]",
"[your advice here]", "[your interpretation here]", "[your routine here]",
}
def _is_placeholder(value: str) -> bool:
return value.strip().lower() in _PLACEHOLDERS or len(value.strip()) < 20
# ── Strip any extra keys the model may add ────────────────────────────────────
def _extract_three_keys(parsed: dict) -> dict:
"""Keep only the 3 required keys. Handle alternate key names the model might use."""
aliases = {
'conseils': ['conseils', 'advice', 'tips', 'recommandations'],
'interpretation': ['interpretation', 'analysis', 'analyse'],
'routine': ['routine', 'daily_routine', 'skincare_routine'],
}
result = {}
for canonical, names in aliases.items():
for name in names:
if name in parsed:
result[canonical] = parsed[name]
break
return result
# ── Prompt builder ─────────────────────────────────────────────────────────────
def _build_prompt(skin_issues: str, language: str) -> str:
"""
Instruction-style prompt with a filled-in example so the model understands
it must generate real content β€” not echo placeholder strings.
Enforces exactly 3 keys.
"""
if language == 'fr':
return (
"Tu es un dermatologue expert. Problèmes de peau détectés par IA chez un patient:\n"
f"{skin_issues}\n\n"
"RΓ©dige une analyse dermatologique professionnelle.\n"
"Réponds UNIQUEMENT avec un objet JSON valide. Pas de texte avant ni après. "
"Exactement 3 clΓ©s, rien d'autre:\n\n"
'"conseils" : conseils pratiques et spécifiques pour traiter ces problèmes précis (2-4 phrases)\n'
'"interpretation" : causes probables, liens entre les problèmes et gravité globale (2-3 phrases)\n'
'"routine" : routine quotidienne complète adaptée à ces problèmes, '
'terminer obligatoirement par la phrase exacte: '
'"Consultez la section des produits recommandΓ©s ci-dessous pour voir les solutions adaptΓ©es Γ  vos besoins spΓ©cifiques."\n\n'
"Exemple de format (remplace le contenu par ta vraie analyse):\n"
'{"conseils":"Utilisez un nettoyant doux contenant de l\'acide salicylique pour traiter les pores dilatés et les cicatrices d\'acné. Évitez de toucher votre visage pour prévenir l\'inflammation.","interpretation":"La présence de pores dilatés et de cicatrices indique une peau à tendance acnéique avec excès de sébum. La gravité est modérée et nécessite une routine adaptée.","routine":"Le matin, nettoyez avec un gel doux puis appliquez un sérum à la niacinamide et une crème solaire SPF 30. Le soir, utilisez un rétinol léger suivi d\'un hydratant. Exfoliez chimiquement 2 fois par semaine. Consultez la section des produits recommandés ci-dessous pour voir les solutions adaptées à vos besoins spécifiques."}'
)
else:
return (
"You are an expert dermatologist. AI-detected skin issues for a patient:\n"
f"{skin_issues}\n\n"
"Write a professional dermatological analysis.\n"
"Reply ONLY with a valid JSON object. No text before or after. "
"Exactly 3 keys, nothing else:\n\n"
'"conseils" : specific practical advice to treat these exact detected issues (2-4 sentences)\n'
'"interpretation" : probable causes, links between issues, and overall severity (2-3 sentences)\n'
'"routine" : complete daily skincare routine tailored to these issues, '
'MUST end with the exact phrase: '
'"Check the recommended products section below to see solutions tailored to your specific needs."\n\n'
"Example format (replace content with your real analysis):\n"
'{"conseils":"Use a gentle salicylic acid cleanser twice daily to address enlarged pores and acne scars. Avoid picking at blemishes to prevent further scarring and inflammation.","interpretation":"The combination of enlarged pores and acne scars suggests an acne-prone, oily skin type with post-inflammatory changes. Overall severity is moderate and manageable with consistent skincare.","routine":"Morning: gentle cleanser, niacinamide serum, SPF 30 sunscreen. Evening: mild retinol followed by a lightweight moisturiser. Exfoliate chemically with AHA or BHA twice a week. Check the recommended products section below to see solutions tailored to your specific needs."}'
)
# ── Core Groq call ─────────────────────────────────────────────────────────────
def _call_groq(detections: list, language: str) -> dict:
client = _get_client()
if not client:
return get_default_analysis(language)
# Deduplicate before building the prompt β€” keeps token count minimal
deduped = _deduplicate(detections)
skin_issues = "\n".join(
f"- {d['class_name']} (confidence: {d['confidence']:.2f})"
for d in deduped
)
print(f"[analysis] Sending {len(deduped)} unique classes to Groq (was {len(detections)} detections)")
prompt = _build_prompt(skin_issues, language)
t0 = time.time()
completion = client.chat.completions.create(
messages=[{"role": "user", "content": prompt}],
model=GROQ_MODEL,
max_tokens=350,
stream=False,
)
print(f"[analysis] Groq response time: {time.time()-t0:.2f}s")
content = completion.choices[0].message.content
match = re.search(r'\{.*\}', content, re.DOTALL)
if match:
try:
parsed = json.loads(match.group(0))
cleaned = _extract_three_keys(parsed)
# Fallback: direct key grab if alias mapping missed them
if not cleaned:
cleaned = {k: parsed[k] for k in ('conseils', 'interpretation', 'routine') if k in parsed}
required = ('conseils', 'interpretation', 'routine')
if all(k in cleaned and not _is_placeholder(str(cleaned[k])) for k in required):
return {k: cleaned[k] for k in required}
print("[analysis] Groq returned incomplete or placeholder values β€” falling back")
except json.JSONDecodeError:
print("[analysis] JSON parse error β€” falling back")
return get_default_analysis(language)
# ── Public entry point (with cache + timeout) ──────────────────────────────────
def analyze_skin_issues(detections: list, language: str = 'en') -> dict:
"""
Returns AI analysis dict with exactly 3 keys: conseils, interpretation, routine.
Hits cache first; falls back to default on timeout or error.
"""
key = _cache_key(detections, language)
with _cache_lock:
if key in _cache:
print("[analysis] Cache hit βœ“")
return _cache[key]
result: dict = {}
exception: Optional[Exception] = None
def worker():
nonlocal exception
try:
result['data'] = _call_groq(detections, language)
except Exception as exc:
exception = exc
t = threading.Thread(target=worker, daemon=True)
t.start()
t.join(GROQ_TIMEOUT)
if t.is_alive():
print("[analysis] Groq timed out β€” returning default")
return get_default_analysis(language, timeout=True)
if exception:
print(f"[analysis] Groq error: {exception}")
return get_default_analysis(language)
analysis = result.get('data', get_default_analysis(language))
_store(key, analysis)
return analysis