Vector search optimization на собеседовании Data Scientist

Зачем разбирать на собесе

Vector search в production требует tuning. На собесе DS: «как scale до миллиардов», «recall budget».

Recall vs latency trade-off

ANN (Approximate Nearest Neighbors) — trade-off: lower recall, faster search.

Recall. Доля true nearest, найденных в approximate top-K. Если true k=10, ANN возвращает 9 of them — recall = 0.9.

Tuning. Каждый ANN method имеет params для recall ↔ latency.

HNSW efSearch: higher → recall+, latency+
IVF nprobe: higher → recall+, latency+

Production target: recall ≥ 0.95-0.99 на target latency.

HNSW параметры

Build time:

• M. Max connections per layer. 16-64 typical. Higher → better recall, more memory.
• efConstruction. Search width при build. 200-500.

Query time:

• efSearch. Search width при query. Tune to target recall.

hnsw.efSearch = 100  # быстрее, recall ниже
hnsw.efSearch = 500  # медленнее, recall выше

IVF + PQ

IVF. Divide vectors на nlist clusters (Voronoi). Search — пробег только nprobe ближайших clusters.

nlist = 4 × √N    (heuristic)
nprobe ≈ 10-100   (tunable)

PQ (Product Quantization). Compress vectors. d=128-dim → split into 8 sub-vectors of 16-dim → kmeans-like compression each.

Original: 128 floats × 4 bytes = 512 bytes per vector
PQ: 8 × 1 byte = 8 bytes per vector (64× less)

Recall падает на 2-5%, memory на 60×.

IVF-PQ. Combine — clustering + compression. На billion-scale — single GPU memory.

Reranking

После ANN retrieve top-K (e.g., 100), rerank с exact similarity на actual vectors.

1. ANN: query → 100 candidates (approximate).
2. Exact dot product / cosine на 100 candidates.
3. Sort, return top-10.

100 exact computations cheap. Final accuracy — exact.

Cross-encoder rerank. Для search. После dense retrieval — heavier model оценивает (query, doc) pairs.

Hybrid search

Dense + Sparse (BM25). Combine semantic + keyword.

final = α · dense_score + (1-α) · bm25_score

Better recall на keyword-heavy queries.

Reciprocal rank fusion (RRF). Combines ranking без weight tuning.

score(d) = Σ_methods 1 / (k + rank_method(d))

Связанные темы

• Vector databases на собесе DS
• Two-tower DSSM для DS
• Cosine vs Euclidean для DS
• Embeddings на собесе DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на документации FAISS / Pinecone / pgvector.

Тренируйте Data Science — откройте тренажёр с 1500+ вопросами для собесов.