qdrant-client

Idiomatic, RAG-oriented Qdrant client for Crystal — a thin, stable wrapper over qdrant-api (the generated, fully-typed transport layer). It exposes exactly the vector working set a retrieval engine needs — create a collection, upsert vectors, KNN search, delete by id, count — and nothing else.

Need the full Qdrant surface (filters, scroll, snapshots, cluster, aliases, sparse/multi-vectors…)? Reach for qdrant-api directly. This shard is the opinionated sugar on top of it.

📖 API documentation

Features

• Stable, hand-written API — Qdrant::Collection + Qdrant::Hit. No generated type ever leaks into the public surface (anti-corruption layer): when qdrant-api is regenerated against a new Qdrant release, the compiler localizes any drift to a handful of internal call sites, not your code.
• The RAG working set, ~5 ops — ensure, upsert (single + batch), search, delete(ids), count. Deliberately no filter DSL: deletes go by id, KNN is unfiltered (fuse on your side).
• Connection as the happy path — an existing, often remote Qdrant (Qdrant Cloud): HTTPS + api-key header, one dedicated collection per corpus.
• Tested against a real Qdrant — the spec suite runs against a Qdrant container, doubling as a contract/canary check in CI.

Installation

Add the dependency to your shard.yml:

dependencies:
  qdrant-client:
    github: jbox-web/qdrant-client.cr

Then run shards install.

Usage

require "qdrant-client"

# An existing Qdrant, often remote / Cloud → HTTPS + api-key header.
docs = Qdrant::Collection.new("docs",
  url: "https://my-qdrant:6333",
  api_key: ENV["QDRANT_API_KEY"]?)

docs.ensure(dim: 768, distance: :cosine)          # idempotent
docs.upsert(42_i64, embedding)                     # single
docs.upsert([                                      # batch
  {43_i64, embedding_a},
  {44_i64, embedding_b},
])

hits = docs.search(query_embedding, top_k: 20)     # => Array(Qdrant::Hit){id, score}
hits.first.id      # Int64
hits.first.score   # Float32

docs.delete([42_i64, 43_i64])                      # purge by id
docs.count                                         # Int64

Qdrant::Hit is the only stable result type — {id : Int64, score : Float32, payload : Hash(String, JSON::Any)}. Search returns {id, score}; hydrate the content from your own store of record by id.

Compatibility

The chain qdrant-client → qdrant-api → pinned OpenAPI spec → Qdrant server isn't obvious from a version number alone, so it's spelled out here:

Qdrant point ids are modeled as Int32 by the OpenAPI layer (ExtendedPointId). Qdrant::Hit#id is exposed as Int64 but round-trips through Int32 — safe as long as ids stay below 2³¹.

Development

mise dev:deps        # shards install
mise dev:qdrant-up   # local Qdrant container (integration specs)
mise dev:check       # build + ameba + specs
mise dev:qdrant-down

License

MIT — Nicolas Rodriguez.