Specs

The shapes referenced by operations. These types are the on-disk vocabulary — they are deliberately decoupled from qdrant-client, so a client-crate upgrade can't silently change the meaning of a committed migration.

CollectionSpec

The full specification of a collection. Used by create_collection and as the desired state passed to revector diff.

VectorSpec

Configuration of a single (dense) named vector.

SparseVectorSpec

Configuration of a single named sparse vector.

VectorParamsDiff

Used inside update_collection.vectors to patch the in-place tunables of an existing named vector. size and distance are deliberately excluded — they are immutable.

HnswConfigSpec

HNSW index parameters. Only fields you set are sent — unset means "leave alone".

QuantizationSpec

Tagged union — set exactly one variant. Use disabled inside update_collection to turn quantization off.

# scalar
quantization_config:
  scalar:
    type: int8
    quantile: 0.99
    memory: pinned      # or the older `always_ram: true`

# product
quantization_config:
  product:
    compression: x8
    always_ram: true

# binary
quantization_config:
  binary:
    always_ram: true

# turboquant
quantization_config:
  turboquant:
    bits: "1.5"
    always_ram: true

# disable (update_collection only)
quantization_config: disabled

ScalarQuantizationSpec

ProductQuantizationSpec

BinaryQuantizationSpec

TurboquantQuantizationSpec

TurboQuant is Qdrant's fast, data-oblivious quantization (Qdrant 1.18+). It encodes each component down to a handful of bits without a training pass.

OptimizersConfigSpec

Optimizer thresholds and behavior. All fields optional.

Distance

Cosine · Euclid · Dot · Manhattan

Datatype

float32 · uint8 · float16 · turbo4 · turbo8 · turbo16

turbo4 stores the primary vectors with TurboQuant 4-bit encoding and requires Qdrant 1.18.2+. turbo8 and turbo16 require Qdrant 1.20+ and store primary vectors at 8 or 16 bits per dimension. Datatypes are selected when creating a collection or adding a vector; changing them requires a new vector.

Memory

cold · cached · pinned — where a component's data is held in RAM (Qdrant 1.19+).

Data is always persisted on disk regardless of this setting — memory only controls caching. It supersedes the older on_disk / always_ram / on_disk_payload booleans, which Qdrant still accepts; when both are set on a component, memory wins. revector sends whichever one a migration declared, so a file written against Qdrant 1.18 keeps its exact meaning.

Qdrant rejects pinned for dense vector storage and for the payload store.

memory needs a Qdrant 1.19 server. Declaring it against an older one would be a silent no-op — gRPC drops fields a server predates rather than rejecting them — so up and down check the live server version first and refuse the run, naming the field and revision. The same guard covers the collection payload: block, params.prefix, params.stemmer: disabled, and datatype: turbo4 (which needs 1.18.2). Qdrant 1.20 settings — id_tracker, turbo8, turbo16, TurboQuant bits: "8", index nested / scoring, and tokenizer: code — are guarded at 1.20.0.

PayloadStorageSpec

Payload storage configuration (Qdrant 1.19+) — the successor to on_disk_payload.

payload:
  memory: cold

IdTrackerSpec

Memory placement of the point ID mapping in indexed segments (Qdrant 1.20+). Accepted on create_collection.spec and update_collection; diff compares placement only when it is declared.

id_tracker:
  memory: cold

TextScoringSpec

Enable BM25 ranking on a full-text payload index (Qdrant 1.20+):

scoring: { type: bm25 }

BM25 statistics are derived from indexed payload text. Applications select the payload field through using in a Qdrant text query; query-time k and b parameters belong to the application, not the index migration.

PayloadSchemaType

keyword · integer · float · geo · text · bool · datetime · uuid

PayloadIndexParamsSpec

Tuning knobs for a payload index, set via create_payload_index.params. Qdrant models these as a per-field-type union; since the field type is already named by schema:, revector keeps the YAML flat and validates applicability instead — setting a field the chosen type doesn't accept fails revector validate offline rather than being silently dropped.

The code tokenizer splits identifiers at underscores and case changes, while also indexing the full identifier. getUserName can match user name or get_user_name.

With nested, the indexed field must be inside the named array. For text, each element becomes its own document for all queries and BM25 scoring. For all types, an is_empty condition on the field matches a point when an element lacks that field, and never a point without elements.

StopwordsSpec

StemmerSpec

stemmer: { snowball: english }   # Snowball stemmer for a language
stemmer: disabled                # opt out of the language default (Qdrant 1.19+)