Les pipelines RAG se cassent lorsque leur unique fournisseur de recherche tombe en panne. Ce tutoriel construit une couche de recherche multi-fournisseur qui interroge Scavio en tant que principal et bascule sur Exa ou Brave en cas d'échec. Le résultat est une disponibilité de recherche proche de 100 % pour votre pipeline RAG, avec un formatage cohérent des résultats quel que soit le fournisseur qui sert la requête.
Prérequis
- Python 3.8+
- bibliothèque requests
- Une clé API Scavio depuis scavio.dev
- Optionnel : clés API Exa et Brave pour le basculement
Parcours
Étape 1: Construire l'interface de recherche unifiée
Créer un client de recherche qui normalise les résultats de plusieurs fournisseurs.
Python
import os, requests, time, json
class UnifiedSearch:
def __init__(self):
self.scavio_key = os.environ.get('SCAVIO_API_KEY', '')
self.exa_key = os.environ.get('EXA_API_KEY', '')
self.brave_key = os.environ.get('BRAVE_API_KEY', '')
self.stats = {'scavio': 0, 'exa': 0, 'brave': 0, 'failures': 0}
def _scavio(self, query, n=5):
resp = requests.post('https://api.scavio.dev/api/v1/search',
headers={'x-api-key': self.scavio_key, 'Content-Type': 'application/json'},
json={'query': query, 'country_code': 'us'}, timeout=10)
resp.raise_for_status()
return [{'title': r.get('title', ''), 'url': r.get('link', ''),
'text': r.get('snippet', ''), 'source': 'scavio'}
for r in resp.json().get('organic_results', [])[:n]]
def _exa(self, query, n=5):
if not self.exa_key: raise Exception('No Exa key')
resp = requests.post('https://api.exa.ai/search',
headers={'x-api-key': self.exa_key, 'Content-Type': 'application/json'},
json={'query': query, 'numResults': n}, timeout=10)
resp.raise_for_status()
return [{'title': r.get('title', ''), 'url': r.get('url', ''),
'text': r.get('text', '')[:200], 'source': 'exa'}
for r in resp.json().get('results', [])[:n]]
def _brave(self, query, n=5):
if not self.brave_key: raise Exception('No Brave key')
resp = requests.get('https://api.search.brave.com/res/v1/web/search',
headers={'X-Subscription-Token': self.brave_key},
params={'q': query}, timeout=10)
resp.raise_for_status()
return [{'title': r.get('title', ''), 'url': r.get('url', ''),
'text': r.get('description', ''), 'source': 'brave'}
for r in resp.json().get('web', {}).get('results', [])[:n]]
def search(self, query, n=5):
for name, fn in [('scavio', self._scavio), ('exa', self._exa), ('brave', self._brave)]:
try:
results = fn(query, n)
self.stats[name] += 1
return {'provider': name, 'results': results, 'count': len(results)}
except Exception as e:
self.stats['failures'] += 1
return {'provider': 'none', 'results': [], 'count': 0}
usearch = UnifiedSearch()
result = usearch.search('rag pipeline best practices 2026')
print(f'Provider: {result["provider"]} | Results: {result["count"]}')
for r in result['results'][:3]:
print(f' [{r["source"]:7}] {r["title"][:50]}')Étape 2: Ajouter un formatage de résultats optimisé pour RAG
Formater les résultats de recherche spécifiquement pour l'injection de contexte RAG.
Python
def format_for_rag(search_result, max_context_chars=3000):
"""Format search results as RAG context with source tracking."""
if not search_result.get('results'):
return {'context': '', 'sources': [], 'char_count': 0}
context_parts = []
sources = []
char_count = 0
for i, r in enumerate(search_result['results']):
source_ref = f'[{i+1}]'
text = r.get('text', '').strip()
if not text:
continue
entry = f'{source_ref} {text}'
if char_count + len(entry) > max_context_chars:
break
context_parts.append(entry)
sources.append({'ref': source_ref, 'title': r['title'][:60], 'url': r['url']})
char_count += len(entry)
context = '\n\n'.join(context_parts)
return {
'context': context,
'sources': sources,
'char_count': char_count,
'provider': search_result.get('provider', 'unknown'),
}
# Build RAG context
result = usearch.search('how to implement vector search python 2026')
rag = format_for_rag(result, max_context_chars=2000)
print(f'\n=== RAG Context ===')
print(f' Provider: {rag["provider"]}')
print(f' Context length: {rag["char_count"]} chars')
print(f' Sources: {len(rag["sources"])}')
for s in rag['sources']:
print(f' {s["ref"]} {s["title"][:50]}')
print(f' {s["url"][:60]}')
print(f'\n Context preview: {rag["context"][:150]}...')Étape 3: Intégrer au pipeline RAG
Brancher la recherche multi-fournisseur dans l'étape de récupération de votre pipeline RAG.
Python
def rag_retrieve(question, max_sources=5):
"""RAG retrieval step using multi-provider search."""
# Primary search
result = usearch.search(question, n=max_sources)
rag_context = format_for_rag(result)
# If primary gives weak results, try refined query
if len(rag_context['sources']) < 2:
refined = usearch.search(f'{question} tutorial guide', n=max_sources)
refined_context = format_for_rag(refined)
if len(refined_context['sources']) > len(rag_context['sources']):
rag_context = refined_context
return rag_context
def rag_pipeline(question):
"""Full RAG pipeline: retrieve, format, generate."""
print(f'\n Question: {question}')
# Step 1: Retrieve
context = rag_retrieve(question)
print(f' Retrieved: {len(context["sources"])} sources via {context["provider"]}')
print(f' Context: {context["char_count"]} chars')
# Step 2: Would pass to LLM here
print(f' Sources for citation:')
for s in context['sources']:
print(f' {s["ref"]} {s["title"][:45]}')
return context
print('=== Multi-Provider RAG Pipeline ===')
for q in ['best vector database 2026', 'how to optimize rag pipeline']:
rag_pipeline(q)
print(f'\n Provider stats: {json.dumps(usearch.stats)}')
print(f' Primary: Scavio $0.005/query')
print(f' Fallback: Exa $0.007/query, Brave ~$0.005/query')
print(f' Uptime target: 99.9% with multi-provider')Exemple Python
Python
import os, requests
SH = {'x-api-key': os.environ['SCAVIO_API_KEY'], 'Content-Type': 'application/json'}
def rag_search(query):
data = requests.post('https://api.scavio.dev/api/v1/search',
headers=SH, json={'query': query, 'country_code': 'us'}, timeout=10).json()
context = '\n'.join(r.get('snippet', '') for r in data.get('organic_results', [])[:3])
sources = [r.get('link', '') for r in data.get('organic_results', [])[:3]]
return {'context': context, 'sources': sources}
result = rag_search('vector database comparison')
print(f'Context: {len(result["context"])} chars, Sources: {len(result["sources"])}')Exemple JavaScript
JavaScript
const SH = { 'x-api-key': process.env.SCAVIO_API_KEY, 'Content-Type': 'application/json' };
const data = await fetch('https://api.scavio.dev/api/v1/search', {
method: 'POST', headers: SH,
body: JSON.stringify({ query: 'vector database comparison', country_code: 'us' })
}).then(r => r.json());
const context = (data.organic_results || []).slice(0, 3).map(r => r.snippet).join('\n');
console.log(`Context: ${context.length} chars`);Sortie attendue
JSON
Provider: scavio | Results: 5
[scavio ] Best Vector Databases for AI in 2026 - Compari
[scavio ] Pinecone vs Weaviate vs Qdrant - Complete Guide
=== RAG Context ===
Provider: scavio
Context length: 1,450 chars
Sources: 5
[1] Best Vector Databases for AI in 2026
https://...
=== Multi-Provider RAG Pipeline ===
Question: best vector database 2026
Retrieved: 5 sources via scavio
Context: 1,450 chars
Provider stats: {"scavio": 4, "exa": 0, "brave": 0, "failures": 0}
Primary: Scavio $0.005/query
Uptime target: 99.9% with multi-provider