How to Evaluate Image Search in Qdrant Using Quepid Part 1 (No Hacks)

Part 1 of a series on building a pipeline to evaluate vector retrieval quality with IR metrics. This installment is data preparation — the search itself is kept intentionally simple. The hacks come in Part 2.


Stack

Data Preparation

  1. Queries — load the shopping-queries dataset into a Pandas DataFrame.
  2. Product selection — from 1M+ US products: drop books, drop products without images, sample 40,000 and download images, keep ~21,589 products with valid image URLs.
  3. Embeddings (Multimodal Embeddings via CLIP, using Qdrant’s fastembed):
    • Images → Qdrant/clip-ViT-B-32-vision → 512-dim vectors
    • Queries → Qdrant/clip-ViT-B-32-text (same space) → enables cross-modal text→image search
  4. Query filtering — keep the top 200 most frequent queries as the evaluation set.

Qdrant Implementation

  • Collection set up via Docker, cosine distance.
  • Products indexed with payloads (title, image URL).
  • Cross-modal validation: the text query “reclining leather sectional” successfully retrieved matching furniture images — “we just found those images by searching with text query.”

Part 2 Preview

Integrating Quepid with this image-search system requires non-standard workarounds — the “cheats and hacks” of the series title.

People

Tools