examples/faiss/retrieval_evals.py

# pip install faiss-cpu datasets sentence-transformers
import numpy as np
import faiss
from datasets import load_dataset
from sentence_transformers import SentenceTransformer

# Configurations
K = 10
MODEL_NAME = 'intfloat/multilingual-e5-small'
LANG = 'en' # pt or fi
INDEX_TYPE = "flat" # or HNSW
USE_PREFIXES = True

dataset_id = "eherra/EuroPIRQ-retrieval"
ds_corpus = load_dataset(dataset_id, f"{LANG}-corpus")['test']
ds_queries = load_dataset(dataset_id, f"{LANG}-queries")['test']
ds_qrels = load_dataset(dataset_id, f"{LANG}-qrels")['test']

model = SentenceTransformer(MODEL_NAME)

corpus_texts = ds_corpus['text']
query_texts = ds_queries['text']

if USE_PREFIXES:
    corpus_texts = [f"passage: {t}" for t in corpus_texts]
    query_texts = [f"query: {t}" for t in query_texts]

corpus_embeddings = model.encode(corpus_texts, normalize_embeddings=True, show_progress_bar=True)
query_embeddings = model.encode(query_texts, normalize_embeddings=True, show_progress_bar=True)

dim = corpus_embeddings.shape[1]
index = faiss.IndexFlatIP(dim) if INDEX_TYPE == "flat" else faiss.IndexHNSWFlat(dim, 32)
index.add(corpus_embeddings)

corpus_ids = np.array([str(curr_id) for curr_id in ds_corpus['id']])

relevant_map = {}
for q_id, c_id in zip(ds_qrels['query-id'], ds_qrels['corpus-id']):
    relevant_map.setdefault(str(q_id), set()).add(str(c_id))

rr_list = []
# Get IDs for queries to match with relevant_map
query_ids = [str(qid) for qid in ds_queries['id']]

for i in range(len(query_ids)):
    q_id = query_ids[i]
    _, indices = index.search(query_embeddings[i:i+1], K)
    
    retrieved_ids = corpus_ids[indices[0]]
    targets = list(relevant_map.get(q_id, set()))
    
    matches = np.where(np.isin(retrieved_ids, targets))[0]
    rr_list.append(1.0 / (matches[0] + 1) if matches.size > 0 else 0.0)

mrr = np.mean(rr_list)

print(f"Model: {MODEL_NAME}")
print(f"Index type: {INDEX_TYPE}")
print(f"Language: {LANG}")
print(f"Mean Reciprocal Rank (MRR@{K}): {mrr:.4f}")