Multilingual semantic search retrieves related content across languages by placing a sentence and its translation near the same point in embedding space. A bi-encoder can encode documents offline, then encode a query online and search the stored vectors without translating the query. section 2.3.2
Cross-language retrieval
A Spanish query such as “zapatos de correr” can reach English content among 100,000 Quora questions without a translation step. On English-Spanish STS 2017 pairs, the multilingual distilled model scores 83.1 Spearman correlation, versus 54.7 for an English-only fine-tuned model. section 2.3.2
Production design separates representation quality from retrieval cost:
- Precompute document embeddings and search them in milliseconds at query time. section 2.3.2
- Do not assume cheap encoding makes comparison cheap: full pairwise similarity remains quadratic, motivating FAISS or batched approximations. section 2.3.2
Listing 2.5 Practical sentence-level similarity with Sentence-BERT.
from sentence_transformers import SentenceTransformer
model = SentenceTransformer('all-MiniLM-L6-v2') #A
tickets = [ #B
"My internet connection keeps dropping",
"WiFi connectivity is unstable",
"Billing question about charges"
]
ticket_embeddings = model.encode(tickets) #C
query = "Network connection problems" #D
query_embedding = model.encode([query])
similarities = model.similarity(query_embedding, ticket_embeddings) #E
print(f"--- Similarities for query: '{query}' ---")
for i, score in enumerate(similarities[0]):
print(f"Score: {score:.3f} | Ticket: '{tickets[i]}'")
A compact Sentence-BERT example shows the baseline workflow: encode texts, encode a query, and compare embeddings before extending the setup across languages. section 2.3.2
