package qdrant.serverless

Get desktop application:
View/edit binary Protocol Buffers messages

CollectionsService manages the collections of a qdrant serverless space. Unlike the qdrant server API, it exposes only a simplified configuration: the manager turns it into concrete qdrant settings, which are never exposed to the tenant.

• rpc CreateCollection (CreateCollectionRequest, CreateCollectionResponse)

  serverless_collections.proto:17

  Creates a collection with the given tenant-facing configuration.

  message CreateCollectionRequest

  serverless_collections.proto:216

  Every request names the collection by its tenant-facing name only. The tenant (`x-account-id`, `x-space-id`) travels in gRPC metadata, injected by auth. The storage id is the manager's: minted on create, resolved internally on get/delete.

  • string collection_name = 1

    Tenant-facing name of the collection to create.

  • optional CollectionConfig config = 2

    The collection's configuration.

  message CreateCollectionResponse

  serverless_collections.proto:224

  Result of a create.

  • string collection_name = 1

    Tenant-facing name of the collection.

  • string result = 2

    Outcome, e.g. "created", "already exists".

  • double time = 3

    Time spent to process

• rpc DeleteCollection (DeleteCollectionRequest, DeleteCollectionResponse)

  serverless_collections.proto:19

  Deletes a collection and all of its data.

  message DeleteCollectionRequest

  serverless_collections.proto:234

  Names the collection to delete.

  • string collection_name = 1

    Tenant-facing name of the collection to delete.

  message DeleteCollectionResponse

  serverless_collections.proto:240

  Result of a delete.

  • bool deleted = 1

    Whether the collection existed and was deleted.

  • double time = 2

    Time spent to process

• rpc GetCollection (GetCollectionRequest, GetCollectionResponse)

  serverless_collections.proto:21

  Returns a single collection's configuration and stats.

  message GetCollectionRequest

  serverless_collections.proto:248

  Names the collection to fetch.

  • string collection_name = 1

    Tenant-facing name of the collection.

  message GetCollectionResponse

  serverless_collections.proto:254

  The collection's configuration and stats, if it exists.

  • bool exists = 1

    Whether the collection exists.

  • optional CollectionConfig config = 2

    The configuration the collection was created with.

  • optional uint64 point_count = 3

    Available points as of the last applied write (eventually consistent); absent until the updater has written stats for the collection.

  • double time = 4

    Time spent to process

• rpc ListCollections (ListCollectionsRequest, ListCollectionsResponse)

  serverless_collections.proto:23

  Lists the caller's collections.

  message ListCollectionsRequest

  serverless_collections.proto:267

  Lists the caller's collections. The tenant travels in metadata.

  • optional uint32 limit = 1

    Maximum number of collections to return. Defaults to 20 and must not exceed 100.

  • optional string offset_token = 2

    Opaque token returned as `next_offset_token` by the previous page. Clients must not interpret this value.

  message ListCollectionsResponse

  serverless_collections.proto:286

  The caller's collections.

  • repeated CollectionSummary collections = 1

    Collections in this page.

  • optional string next_offset_token = 2

    Opaque token to pass as `offset_token` to retrieve the next page. Absent when there are no more results.

  • double time = 3

    Time spent to process

Exact match on booleans, e.g. `is_active: true`.

Used in: PayloadIndexConfig

(message has no fields)

The tenant's collection config. Persisted verbatim. At least one dense or sparse vector is required.

Used in: CreateCollectionRequest, GetCollectionResponse

• map<string, DenseVectorConfig> dense_vectors = 1

  Keyed by vector name. The empty name "" is the unnamed default vector, as in qdrant; a collection either has the single unnamed vector or named ones.

• map<string, SparseVectorConfig> sparse_vectors = 2

  Keyed by vector name.

• map<string, PayloadIndexConfig> payload_indexes = 3

  Keyed by payload field name (JSON path, e.g. `user_id` or `meta.tags`).

One collection in a listing.

Used in: ListCollectionsResponse

• string collection_name = 1

  Tenant-facing name of the collection.

• optional uint64 point_count = 2

  Available points as of the last applied write (eventually consistent); absent until the updater has written stats for the collection.

Range filters on RFC 3339 datetimes, e.g. `created_at: "2023-02-08T10:49:00Z"`.

Used in: PayloadIndexConfig

(message has no fields)

Configuration of a single dense (embedding) vector.

Used in: CollectionConfig

• uint64 size = 1

  Dimensionality of the embedding, e.g. 512 (CLIP), 1536, 3072.

• Distance distance = 2

  Distance metric used to compare vectors.

• bool multivector = 3

  Store several sub-vectors per point and compare with max-sim, for late-interaction models (ColBERT, ColPali, ...).

• optional PrecisionTier precision_tier = 4

  Precision/cost trade-off for this vector. Unset: HIGH.

Explicitly disable stemming (overrides any language default).

Used in: StemmingAlgorithm

(message has no fields)

Distance metric used to compare dense vectors.

Used in: DenseVectorConfig

• DISTANCE_UNSPECIFIED = 0

  Unset; a concrete metric is required.

• COSINE = 1

  Cosine similarity.

• EUCLID = 2

  Euclidean (L2) distance.

• DOT = 3

  Dot product.

• MANHATTAN = 4

  Manhattan (L1) distance.

Range filters on floating point (and integer) numbers, e.g. `price: 99.5`.

Used in: PayloadIndexConfig

(message has no fields)

Geo radius / bounding box / polygon filters on `{lon, lat}` values.

Used in: PayloadIndexConfig

(message has no fields)

Exact match and/or range filters on integers, e.g. `age: 25`. Both are on by default; turning one off shrinks the index.

Used in: PayloadIndexConfig

• optional bool lookup = 1

  Support exact-match filters.

• optional bool range = 2

  Support range filters.

Exact match on string values, e.g. `color: "red"`.

Used in: PayloadIndexConfig

• optional KeywordPrefixParams prefix = 1

  If set, enable prefix matching (`match: { "prefix": ... }`) on this field. Presence of this message enables prefix matching; it has no options yet.

Prefix matching options for the keyword index. Has no options yet: presence of this message enables prefix matching.

Used in: KeywordIndex

(message has no fields)

One payload index. Only the *kind* of filter the field supports is chosen here; storage placement of the index is the manager's decision.

Used in: CollectionConfig

• oneof index

  The kind of filter the field supports.

  • KeywordIndex keyword = 1

    Exact match on strings.

  • IntegerIndex integer = 2

    Exact match and/or range filters on integers.

  • FloatIndex float = 3

    Range filters on numbers.

  • UuidIndex uuid = 4

    Exact match on UUIDs.

  • DatetimeIndex datetime = 5

    Range filters on datetimes.

  • TextIndex text = 6

    Full-text filtering.

  • GeoIndex geo = 7

    Geo filters.

  • BoolIndex bool = 8

    Exact match on booleans.

How much of the original vector precision may be traded for cost. The manager turns this into a concrete quantization / datatype choice; the tenant never picks scalar/product/binary quantization or its parameters directly.

Used in: DenseVectorConfig, SparseVectorConfig

• PRECISION_TIER_UNSPECIFIED = 0

  Unset: HIGH.

• LOW = 1

  Aggressive compression, lowest cost, approximate results.

• MEDIUM = 2

  Moderate compression with a small accuracy trade-off.

• HIGH = 3

  Light compression, near-exact results.

• FULL = 4

  No lossy compression: exact vectors, fully in memory. Highest cost.

Snowball stemming for a full-text index.

Used in: StemmingAlgorithm

• string language = 1

  Language for the snowball algorithm, e.g. "english".

Configuration of a single sparse vector.

Used in: CollectionConfig

• bool use_idf = 1

  Apply the IDF modifier at query time. Enable for BM25-style models that expect inverse-document-frequency weighting.

• optional PrecisionTier precision_tier = 2

  Precision/cost trade-off for this vector. Unset: HIGH.

Stemming algorithm for a full-text index. Unset: no stemming.

Used in: TextIndex

• oneof stemming_params

  Which stemming algorithm to use.

  • SnowballParams snowball = 1

    Snowball stemmer for the given language.

  • DisabledStemmer disabled = 2

    Explicitly disable stemming.

Tokens ignored by a full-text index. Language names match qdrant (e.g. "english"); predefined lists and custom tokens are merged.

Used in: TextIndex

• repeated string languages = 1

  Languages whose predefined stopword lists to apply.

• repeated string custom = 2

  Extra stopwords to ignore, merged with the language lists.

Full-text filtering on string values.

Used in: PayloadIndexConfig

• optional Tokenizer tokenizer = 1

  Tokenizer to split text with. Unset: WHITESPACE.

• optional bool lowercase = 2

  Lowercase text before indexing. Default true.

• optional bool phrase_matching = 3

  Support phrase queries; extra index structure. Default true.

• optional uint64 min_token_len = 4

  Minimum token length to index.

• optional uint64 max_token_len = 5

  Maximum token length to index.

• optional bool ascii_folding = 6

  Fold accented characters to ASCII. Default false.

• optional StopwordsSet stopwords = 7

  Tokens to ignore at index and query time.

• optional StemmingAlgorithm stemmer = 8

  Stemming algorithm. Unset: engine default (no stemming).

Full-text tokenizer, mirrors qdrant's `TokenizerType`.

Used in: TextIndex

• TOKENIZER_UNSPECIFIED = 0

  Unset: WHITESPACE.

• PREFIX = 1

  Index every prefix of each token.

• WHITESPACE = 2

  Split on whitespace.

• WORD = 3

  Split on word boundaries.

• MULTILINGUAL = 4

  Language-aware tokenization.

Exact match on UUID strings; like keyword but stored compactly.

Used in: PayloadIndexConfig

(message has no fields)