Comment construire un pipeline RAG avec repli de recherche en direct

def score_retrieval_quality(query: str, documents: list, scores: list = None) -> float:
    """Score retrieval quality from 0 (terrible) to 1 (excellent)."""
    if not documents:
        return 0.0
    # If vector store provides similarity scores, use them
    if scores:
        avg_score = sum(scores) / len(scores)
        return min(avg_score, 1.0)
    # Heuristic: check keyword overlap between query and docs
    query_words = set(query.lower().split())
    total_overlap = 0
    for doc in documents:
        doc_words = set(doc.lower().split()[:200])  # first 200 words
        overlap = len(query_words & doc_words) / max(len(query_words), 1)
        total_overlap += overlap
    avg_overlap = total_overlap / len(documents)
    return min(avg_overlap, 1.0)

# Example:
score = score_retrieval_quality(
    'latest python version 2026',
    ['Python 3.12 was released in October 2023 with improved performance.']
)
print(f'Retrieval confidence: {score:.2f}')  # Low because doc is outdated

import requests, os

API_KEY = os.environ['SCAVIO_API_KEY']

def search_fallback(query: str, k: int = 5) -> list:
    """Fetch live search results as fallback documents."""
    resp = requests.post('https://api.scavio.dev/api/v1/search',
        headers={'x-api-key': API_KEY, 'Content-Type': 'application/json'},
        json={'query': query, 'country_code': 'us'})
    resp.raise_for_status()
    results = resp.json().get('organic_results', [])[:k]
    return [{
        'content': f'{r["title"]}\n{r.get("snippet", "")}',
        'source': r['link'],
        'retriever': 'live_search'
    } for r in results]

# Test:
fallback_docs = search_fallback('latest python version 2026')
for doc in fallback_docs[:2]:
    print(f'[{doc["retriever"]}] {doc["content"][:80]}...')

CONFIDENCE_THRESHOLD = 0.3  # below this, trigger fallback

def retrieve_with_fallback(query: str, vector_store, k: int = 5) -> dict:
    """Retrieve from vector store; fall back to search if confidence is low."""
    # Step 1: Try vector retrieval
    vector_results = vector_store.similarity_search_with_score(query, k=k)
    docs = [doc.page_content for doc, score in vector_results]
    scores = [score for doc, score in vector_results]
    confidence = score_retrieval_quality(query, docs, scores)
    # Step 2: Decide retrieval strategy
    if confidence >= CONFIDENCE_THRESHOLD and docs:
        return {
            'documents': [{'content': d, 'retriever': 'vector'} for d in docs],
            'strategy': 'vector',
            'confidence': round(confidence, 3),
            'search_cost': 0
        }
    # Step 3: Fallback to live search
    search_docs = search_fallback(query, k=k)
    return {
        'documents': search_docs + [{'content': d, 'retriever': 'vector'} for d in docs[:2]],
        'strategy': 'search_fallback',
        'confidence': round(confidence, 3),
        'search_cost': 0.005
    }

def rag_with_fallback(query: str, vector_store, llm) -> dict:
    # Retrieve with fallback
    retrieval = retrieve_with_fallback(query, vector_store)
    documents = retrieval['documents']
    # Build context
    context = '\n\n'.join(d['content'] for d in documents)
    sources = [d.get('source', 'vector store') for d in documents if d.get('source')]
    # Generate answer
    prompt = f"""Answer based on the following context. If using web results, cite the URLs.

Context:
{context}

Question: {query}
Answer:"""
    # Assuming llm is a callable that returns text
    answer = llm(prompt)
    return {
        'answer': answer,
        'strategy': retrieval['strategy'],
        'confidence': retrieval['confidence'],
        'sources': sources,
        'cost': retrieval['search_cost']
    }

# Usage stays the same as before:
# result = rag_with_fallback(user_query, my_vector_store, my_llm)
# print(result['answer'])
# print(f'Strategy: {result["strategy"]}, Cost: ${result["cost"]}')

from collections import defaultdict

fallback_stats = defaultdict(int)

def tracked_rag(query: str, vector_store, llm) -> dict:
    result = rag_with_fallback(query, vector_store, llm)
    fallback_stats['total'] += 1
    fallback_stats[result['strategy']] += 1
    fallback_stats['total_cost'] += result['cost']
    return result

def print_fallback_report():
    total = fallback_stats['total']
    if total == 0:
        print('No queries tracked yet.')
        return
    vector_pct = fallback_stats.get('vector', 0) / total * 100
    fallback_pct = fallback_stats.get('search_fallback', 0) / total * 100
    print(f'RAG Fallback Report:')
    print(f'  Total queries: {total}')
    print(f'  Vector retrieval: {vector_pct:.0f}%')
    print(f'  Search fallback: {fallback_pct:.0f}%')
    print(f'  Total search cost: ${fallback_stats["total_cost"]:.2f}')
    print(f'  Avg cost/query: ${fallback_stats["total_cost"] / total:.4f}')
    if fallback_pct > 30:
        print(f'  NOTE: High fallback rate. Consider adding more documents to your vector store.')

# After running many queries:
# print_fallback_report()