← All projects

3 · 2026

Aurum Market, semantic catalogue search

A product discovery engine over 15,000 Spanish-language products, built on Qdrant and E5 embeddings. It searches by intent and catches duplicate listings, with no LLM at runtime.

0.92
MRR@10, the first hit almost always comes first
1.0
ANN index fidelity against exact search
59 ms
p50 latency on CPU
379
automated tests
ArchitectureOne index serves both search and duplicate detection, and an exact oracle measures whether the index drops candidates.
  1. Data

    • 15,000-product catalogue10 files verified by SHA-256
  2. Representation

    • Text composition3 strategies compared (E0–E3)
    • e5-base embeddings768 dims, normalised, cached
  3. Store

    • Qdrantapproximate HNSW, what ships
    • NumPy oracleexact search, the yardstick
  1. Retrieval

    • Dense searchbrand filter inside the engine
    • TF-IDFbaseline
  2. Applications

    • Duplicatescalibrated, frozen threshold
    • Events24 ordered mutations
  3. Evaluation

    • MetricsnDCG, Recall, MRR
    • Fidelity and attributionengine vs oracle, which layer fails
  4. Delivery

    • Artefactsvalidated against the contract before writing

Problem

A catalogue of 15,000 Spanish-language products raises two business problems. The first is finding products by intent (“a gift for a cyclist”) rather than exact keywords, sometimes within a specific brand. The second is preventing duplicate products from being listed. Both are solved with a single vector database.

Approach

  • Embeddings with intfloat/multilingual-e5-base (768 dimensions) on CPU, cached by model, role and text digest. The first ingestion takes about 13 minutes; later runs take around 1.
  • Qdrant as the engine, with explicit HNSW parameters (m=24, ef_construct=120, ef_search=256). The brand filter is applied inside the query.
  • An exact NumPy oracle with the same interface as Qdrant, to measure whether the approximate index drops candidates.
  • An experiment matrix (E0–E3) to choose which text of each product becomes the vector.
  • A duplicate threshold calibrated on the development set and frozen before being applied to the blind set.
  • A CLI (aurum doctor, ingest, evaluate, deliver…) that regenerates every deliverable with one command.

I built it with Spec Driven Development: 29 requirements, 8 ADRs recording the rejected alternatives, and no requirement closed without executable evidence.

Results

Metric Value
nDCG@10 0.5422
Recall@10 0.2481, against a ceiling of 0.5188 (48 % of what is achievable)
MRR@10 0.9167
ANN fidelity 1.0000
p50 / p95 latency 59 / 85 ms
Duplicates P / R / F1 1.0 / 1.0 / 1.0 on dev, threshold 0.9191
Brand-filtered queries 10/10, no intruders

What I learned

A metric without context is misleading. With around 25 relevant products per query and only 10 slots, no system can exceed a Recall@10 of 0.5188, so 0.25 looks like a failure until you compare it with that ceiling.

An index fidelity of 1.0 has a strong implication: all the error comes from the representation, not the index, and that is proven with an attribution analysis rather than assumed.

And a detail that is hard to find: Qdrant’s indexing threshold is measured in kilobytes per segment, not in number of vectors. With the defaults, Qdrant silently falls back to brute force and any fidelity measurement comes out trivially perfect.