Les LLM locaux fonctionnant sur llama.cpp, Ollama ou vLLM sont puissants mais figés dans le temps. Ils hallucinent sur l'actualité, les sorties récentes et les données en direct car leurs données d'entraînement ont une date limite. L'ajout d'une API de recherche leur donne un ancrage en temps réel : avant de répondre, le LLM interroge le web et utilise les résultats frais comme contexte. Ce tutoriel fonctionne avec tout endpoint LLM local compatible OpenAI. Coût : 0,005 $ par réponse ancrée.
Prérequis
- Un LLM local en fonctionnement (Ollama, serveur llama.cpp ou vLLM)
- Python 3.9+ installé
- Bibliothèque requests installée
- Une clé API Scavio depuis scavio.dev
Parcours
Étape 1: Connectez-vous à votre LLM local
Configurez la connexion à votre LLM local. Fonctionne avec tout endpoint compatible OpenAI (Ollama, serveur llama.cpp, vLLM).
Python
import requests
# Common local LLM endpoints:
# Ollama: http://localhost:11434/v1/chat/completions
# llama.cpp: http://localhost:8080/v1/chat/completions
# vLLM: http://localhost:8000/v1/chat/completions
LLM_URL = 'http://localhost:11434/v1/chat/completions' # Ollama default
LLM_MODEL = 'llama3' # or 'mistral', 'codellama', etc.
def ask_llm(messages: list, max_tokens: int = 512) -> str:
resp = requests.post(LLM_URL, json={
'model': LLM_MODEL,
'messages': messages,
'max_tokens': max_tokens,
'temperature': 0.3
}, timeout=120)
return resp.json()['choices'][0]['message']['content']
# Test connection
try:
answer = ask_llm([{'role': 'user', 'content': 'Say hello in one word.'}], max_tokens=10)
print(f'LLM connected: {answer}')
except Exception as e:
print(f'LLM connection error: {e}')
print('Make sure Ollama/llama.cpp is running.')Étape 2: Ajoutez la fonction d'ancrage par recherche
Créez une fonction qui interroge le web et formate les résultats comme contexte pour le LLM. Le LLM ne voit que les extraits de recherche, pas les pages entières.
Python
import os
SCAVIO_KEY = os.environ['SCAVIO_API_KEY']
def search_context(query: str, count: int = 5) -> str:
"""Search the web and return formatted context for the LLM."""
resp = requests.post('https://api.scavio.dev/api/v1/search',
headers={'x-api-key': SCAVIO_KEY, 'Content-Type': 'application/json'},
json={'query': query, 'country_code': 'us', 'num_results': count})
results = resp.json().get('organic_results', [])
if not results:
return 'No search results found.'
context = 'Search results (use these to answer accurately):\n\n'
for i, r in enumerate(results, 1):
context += f'[{i}] {r["title"]}\n'
context += f' {r.get("snippet", "")}\n'
context += f' Source: {r["link"]}\n\n'
return context
# Test
ctx = search_context('Python 3.14 release date')
print(ctx[:300])Étape 3: Construisez le pipeline de réponse ancrée
Combinez la recherche et le LLM en une seule fonction. Le LLM reçoit le contexte de recherche et doit citer les sources dans sa réponse.
Python
def grounded_answer(question: str) -> dict:
"""Answer a question using search-grounded local LLM."""
# Step 1: Search for context
context = search_context(question, count=5)
# Step 2: Ask LLM with context
messages = [
{'role': 'system', 'content': (
'You are a helpful assistant. Answer ONLY based on the search results provided. '
'Cite sources as [1], [2], etc. If the search results do not contain the answer, '
'say "I could not find this information in the search results."'
)},
{'role': 'user', 'content': f'{context}\nQuestion: {question}'}
]
answer = ask_llm(messages, max_tokens=512)
return {
'question': question,
'answer': answer,
'grounded': True,
'search_cost': 0.005
}
# Test with a question that requires current data
result = grounded_answer('What is the latest version of Python?')
print(f'Q: {result["question"]}')
print(f'A: {result["answer"]}')
print(f'Grounded: {result["grounded"]}, Cost: ${result["search_cost"]}')Étape 4: Ajoutez un ancrage intelligent (rechercher uniquement quand nécessaire)
Toutes les questions n'ont pas besoin de recherche. Ajoutez une vérification qui décide d'ancrer avec la recherche ou de répondre directement, économisant ainsi des coûts.
Python
def needs_grounding(question: str) -> bool:
"""Heuristic: does this question need real-time data?"""
grounding_triggers = [
'latest', 'current', 'today', '2026', '2025', 'now',
'price', 'cost', 'version', 'release', 'new', 'update',
'best', 'top', 'compare', 'vs', 'alternative',
'how much', 'where to', 'who is',
]
q_lower = question.lower()
return any(trigger in q_lower for trigger in grounding_triggers)
def smart_answer(question: str) -> dict:
"""Answer with search grounding only when needed."""
if needs_grounding(question):
return grounded_answer(question)
# Direct LLM answer (no search cost)
messages = [{'role': 'user', 'content': question}]
answer = ask_llm(messages, max_tokens=512)
return {
'question': question,
'answer': answer,
'grounded': False,
'search_cost': 0
}
# Test both paths
for q in ['What is a Python list comprehension?',
'What is the latest Python version in 2026?']:
result = smart_answer(q)
print(f'[{"GROUNDED" if result["grounded"] else "DIRECT"}] '
f'${result["search_cost"]} - {q}')
print(f' {result["answer"][:100]}...')
print()Exemple Python
Python
import requests, os
LLM_URL = 'http://localhost:11434/v1/chat/completions'
SCAVIO_KEY = os.environ['SCAVIO_API_KEY']
def search(query, count=5):
resp = requests.post('https://api.scavio.dev/api/v1/search',
headers={'x-api-key': SCAVIO_KEY, 'Content-Type': 'application/json'},
json={'query': query, 'country_code': 'us', 'num_results': count})
return resp.json().get('organic_results', [])
def grounded_ask(question):
results = search(question)
ctx = '\n'.join(f'[{i+1}] {r["title"]}: {r.get("snippet","")}' for i, r in enumerate(results))
resp = requests.post(LLM_URL, json={'model': 'llama3', 'messages': [
{'role': 'system', 'content': 'Answer from search results. Cite [1],[2].'},
{'role': 'user', 'content': f'{ctx}\n\nQ: {question}'}], 'max_tokens': 512})
return resp.json()['choices'][0]['message']['content']
print(grounded_ask('latest Python version 2026'))Exemple JavaScript
JavaScript
const LLM_URL = 'http://localhost:11434/v1/chat/completions';
const SCAVIO_KEY = process.env.SCAVIO_API_KEY;
async function groundedAsk(question) {
const searchResp = await fetch('https://api.scavio.dev/api/v1/search', {
method: 'POST',
headers: { 'x-api-key': SCAVIO_KEY, 'Content-Type': 'application/json' },
body: JSON.stringify({ query: question, country_code: 'us', num_results: 5 })
});
const results = (await searchResp.json()).organic_results || [];
const ctx = results.map((r, i) => `[${i+1}] ${r.title}: ${r.snippet || ''}`).join('\n');
const llmResp = await fetch(LLM_URL, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({ model: 'llama3', messages: [
{ role: 'system', content: 'Answer from search results. Cite [1],[2].' },
{ role: 'user', content: `${ctx}\n\nQ: ${question}` }], max_tokens: 512 })
});
return (await llmResp.json()).choices[0].message.content;
}
groundedAsk('latest Python version 2026').then(console.log);Sortie attendue
JSON
LLM connected: Hello
Search results (use these to answer accurately):
[1] Python Release Python 3.14.0
Python 3.14.0 was released on October 7, 2025...
Source: https://www.python.org/downloads/release/python-3140/
Q: What is the latest version of Python?
A: According to the search results, the latest version of Python is 3.14.0,
released on October 7, 2025 [1].
[DIRECT] $0 - What is a Python list comprehension?
A list comprehension is a concise way to create lists...
[GROUNDED] $0.005 - What is the latest Python version in 2026?
The latest Python version is 3.14.0, released October 2025 [1]...