Customers

In production

Three teams, three very different corpora, one shared problem: the index had become the thing they were maintaining instead of the product.

E-commerce Ardent Supply

Search that survives the way people actually type

The problem

A catalogue of 4.2 million parts where half the queries used a phrasing no synonym list covered. The lexical index returned nothing, and “no results” pages were the second most visited screen on the site.

What changed

Hybrid retrieval — BM25 for exact part numbers, vectors for everything else. Empty result pages fell to a rounding error.

61%: fewer zero-result queries
12 ms: p99 at 4.2M vectors

We had spent two years maintaining synonym files. We deleted them in an afternoon and search got better.
Search lead, Ardent Supply

Legal tech Corpus Legal

Retrieval a lawyer will sign their name under

The problem

A RAG pipeline over 900,000 case documents where recall mattered more than latency. Missing one relevant precedent was the only failure mode that counted.

What changed

Tuned efSearch per query class, with metadata filtering on jurisdiction and date applied inside the index rather than after it.

0.98: recall@20, measured
3×: fewer reranking calls

Filtering inside the index instead of after it was the whole difference. Our recall stopped depending on how narrow the filter was.
Head of engineering, Corpus

Developer tools Tessellate

From a weekend prototype to a billion vectors

The problem

A code search product that outgrew pgvector at around 40 million embeddings. Reindexing needed a maintenance window nobody could schedule.

What changed

Migrated in stages behind a read alias, with live reindexing. No downtime, and the on-call rotation stopped carrying the index.

1.1B: vectors in production
0: maintenance windows since

The migration was boring, which is the highest compliment I can pay an infrastructure change.
Co-founder, Tessellate

Patterns

What people actually build

RAG retrieval

Grounding an assistant in a corpus the model never saw at training time.

Product search

Catalogues where intent rarely matches the vocabulary of the listing.

Deduplication

Finding near-identical records across systems that never agreed on a key.

Recommendation

Neighbourhood queries over behavioural embeddings, refreshed continuously.

Your turn

Tell us what you are retrieving

We will say whether the shape of your corpus suits the engine, and what the index would cost to run.

Start free Talk to an engineer