How to Evaluate Image Search in Qdrant Using Quepid Part 2 (Hacks)

The payoff of the series: actually measuring image-search quality. Andrew Kornilov registers Qdrant Vector DB as a custom Quepid search endpoint, loads 200 CLIP queries with embeddings, and computes NDCG@10 — plus the hacks (and one upstream PR) needed to make image results judgeable. Builds on Part 1 and the case-file trick from Oops, I Did It Again.

---

Infrastructure (three containers)

1. Qdrant — REST :6333, gRPC :6334; embeddings + image metadata.
2. Quepid — MySQL 8.4.4 backend + Quepid v8.3.6 app (:3000); judgment management and IR metrics.
3. Quepid HTTP API wrapper — the author’s unofficial API (frutik777/quepid-api-unofficial, :8081) for programmatic case/query creation.

Init steps: DB migrations, seed data, admin user, API key. Python deps: pandas, requests, tqdm.

Building the Evaluation Case

1. Team — created via the API to group evaluations.
2. Search endpoint — register Qdrant as a custom backend with a JavaScript mapper that turns Qdrant responses into Quepid’s format (doc IDs, thumbnails, titles from the payload). The key trick: inject the query embedding from options with "vector": "#$qOption.clip##".
3. Cases & queries — load the 200 pre-selected queries with their CLIP embeddings; Quepid computes NDCG@10 as judges rate.

Image-as-Query Case

A separate use case: sample 25 random products and use their image vectors as queries — image→image visual-similarity evaluation, in its own case.

The Hacks

• Hack 1 — scorer-level rendering. Inject JavaScript into Quepid’s scorer to render image URLs as actual <img> elements: const customHTML = \`;` Limits: only the first results page; breaks pagination; no page control.
• Hack 2 — fix it properly upstream. Rather than maintain fragile JS, the author opened Quepid PR #1683 to add native image-query rendering.

Achievements

1. Qdrant configured as a Quepid endpoint
2. A case of 200 text queries with embeddings
3. Query-construction + result-parsing infrastructure
4. Explored UI customization for image visualization
5. Documented platform limitations needing native fixes

Key Insight

Creative workarounds can extend a platform, but sustainable solutions mean contributing fixes upstream rather than nursing JavaScript hacks.

Future Direction

Next: automated judgement generation via Quepid’s LLM-as-a-judge (LLM as Judge), to cut manual annotation.

Related Concepts

• Vector Search Evaluation — the realized pipeline
• Multimodal Embeddings — CLIP text/image vectors; image→image search
• Quepid · Judgment Lists · NDCG · Search Evaluation
• LLM as Judge — teased follow-up

Related Articles

• How to Evaluate Image Search in Qdrant Using Quepid Part 1 — data prep
• Oops, I Did It Again — the case-file / query-options technique
• Why Setting Up Quepid for Vector Search Evaluation Went Wrong — original blockers
• Creating Judgement Lists with Quepid — the conventional (text) workflow

People

• Andrew Kornilov — author

Tools

• Qdrant Vector DB · Quepid