The first real feature work on engine was fixing something that looked correct but wasn’t.

The search endpoint was filtering by domain after Qdrant returned results. At small corpus size — a handful of fixture entries — it’s invisible. But at scale it’s wrong: Qdrant returns the top-N entries by vector similarity, then Java discards those that don’t match the requested domain. If the top 8 results happen to be from the wrong domain, the caller gets nothing back. No error, no warning, just zero results.

The fix is LangChain4j’s Filter API. Pass an IsEqualTo("domain", domain) to EmbeddingSearchRequest.builder().filter() and the filter travels server-side — Qdrant applies it before scoring, so only matching entries compete for the limit slots.

Filter domainFilter = (domain != null && !domain.isBlank())
        ? new IsEqualTo("domain", domain)
        : null;

var request = EmbeddingSearchRequest.builder()
        .queryEmbedding(queryEmbedding)
        .maxResults(maxResults)
        .filter(domainFilter)
        .build();

I knew Qdrant had payload filtering — I didn’t know LangChain4j exposed it through EmbeddingSearchRequest. We verified the chain from bytecode: IsEqualTo → QdrantFilterConverter.buildEqCondition() → gRPC FieldCondition. Not in the docs but clearly supported.

Writing the tests surfaced a different problem. REST Assured’s Groovy GPath $.findAll { } returns null when no elements match — not []. So hasSize(0) throws even when the filter is working correctly. The fix is anyOf(nullValue(), hasSize(0)). Small thing, but the failure message (“was null”) reads like the assertion found a problem, not that the predicate returned nothing.

Both findings went into the knowledge garden — the GPath null return as a gotcha, the LangChain4j filter technique as a technique. Both would have cost time to discover independently.

<
Previous Post
Neural Text — The Native Image Gate
>
Next Post
Reviewing your way to fewer surprises