ChromaDotNet.Client conformance: Chroma single server and Chroma Cloud

• Client: ChromaDotNet.Client 2.9.2
• Chroma: chromadb/chroma:1.5.9, a single server in Docker
• Chroma Cloud: AWS us-east-1, run on 2026-10-06 with the client 2.9.2 on .NET 10.0.12; every collection and database deleted at the end
• OpenTelemetry SDK: 1.19.1; ChromaDotNet.Testcontainers: 0.1.3 (Testcontainers 4.15.0)
• Chroma image digest: chromadb/chroma@sha256:1e0b73a187a28757c572acba508c46f48c9e8b0acaf5c20e6d95cdedce1acdf6
• Docker: 29.7.2 (linux/amd64)
• .NET SDK: 10.0.400
• Host: Windows 10.0.26200, x86_64
• Client package: from nuget.org
• Arm64: the Linux arm64 images of .NET, emulated by Docker on this x86-64 host
• Environments: 31, 1500 checks run
• Generated: 2026-10-06 08:15 UTC

✓ every check passed · ✗ a check failed · — not on a single server: the Chroma docs put it in Chroma Cloud, and the check verified that the client reports the error of the server as documented · on Chroma Cloud, — is what an API key cannot do, with the error the check verified · ¹ a note under the table · a blank cell: the row is not about that column (the quotas of Chroma Cloud on a single server; on Chroma Cloud, the checks with no server or with a local container)

By feature

¹ On cloud_aws-us-east-1, “create and get a tenant” is not available: an API key of Chroma Cloud cannot create a tenant: Chroma Cloud answers 403 "Permission denied.", which the client throws as a ChromaException; on cloud_aws-us-east-1, “collections in another tenant and database” is not available: an API key of Chroma Cloud cannot create a tenant for the collections of another tenant: Chroma Cloud answers 403 "Permission denied.", which the client throws as a ChromaException.
³ On win_netcoreapp3.1, win_netcoreapp3.1_ja-JP, “the duration histogram and its buckets, with the OpenTelemetry SDK” is not available: the OpenTelemetry SDK exports no metric on this runtime, also from a meter of its own.

Same answers in every environment

The seeded fuzz of the test bench sends the same random requests from each environment to a fresh Chroma 1.5.9 and compares what comes back, suite by suite, with the answers on win_net10.0: 30 environments, two seeds. ✓ the same answers in every environment of the column, ✗ different somewhere.

Environments

Checks

Clients: local server, Chroma Cloud, authentication

• ✓ the address alone, with /api/v2 and with /api/v2/
• ✓ constructors: uri, options, options and HttpClient
• ✓ connection strings
• ✓ version, heartbeat, healthcheck, pre-flight checks, identity
• ✓ authentication headers: X-Chroma-Token, Bearer, Basic
• ✓ the v1 API on Chroma 1.x answers 410 Gone
• ✓ errors: status, error type, CollectionExistsAsync

Tenants and databases

• ✓ create and get a tenant; — on cloud_aws-us-east-1: an API key of Chroma Cloud cannot create a tenant: Chroma Cloud answers 403 "Permission denied.", which the client throws as a ChromaException
• ✓ databases: create, get, list with paging, delete
• ✓ collections in another tenant and database; — on cloud_aws-us-east-1: an API key of Chroma Cloud cannot create a tenant for the collections of another tenant: Chroma Cloud answers 403 "Permission denied.", which the client throws as a ChromaException

Manage collections

• ✓ create, get, get or create, exists, delete
• ✓ list with paging, count
• ✓ modify name and metadata
• ✓ by id, collection clients, dimension, version
• ✓ names and metadata the client stops: dots, lists, lone surrogates

Add, update, delete data (large writes in batches)

• ✓ add and read back embeddings, metadata, documents, URIs
• ✓ update, upsert, null deletes a key
• ✓ 12,000 records: written in batches, read in pages
• ✓ batches: one request, and a smaller size
• ✓ delete: by ids, by filters, with a limit
• ✓ count at a read level

Configure: HNSW, SPANN, embedding function

• ✓ space: l2, cosine, inner product
• ✓ HNSW settings, changed after the creation
• ✓ SPANN on a single server: the client tells
• ✓ embedding function declared
• ✓ schema: value indexes off, full-text search off, space
• ✓ index settings, changed after the creation: SPANN on Chroma Cloud
• ✓ the index Chroma Cloud does not use: HNSW, the client tells

Query and get: l2, cosine, ip

• ✓ nearest records and distances, in each space
• ✓ several queries, ids, filters, include, ChromaQuery
• ✓ get: one, many, missing, limit and offset, include
• ✓ floats the server cannot write: read as NaN or infinity, the other records intact

Metadata filtering, array metadata

• ✓ every operator on ints, floats, strings, bools
• ✓ & and |, nested, long chains
• ✓ lists in metadata: $contains, $not_contains
• ✓ metadata values read exactly, or inferred

Full-text search and regex

• ✓ $contains, $not_contains, combined
• ✓ $regex, $not_regex

Collection forking

• ✓ ForkAsync and ForkCountAsync; — on alpine_net10.0, alpine_net8.0, aot_net10.0, arm64_net10.0, arm64_net8.0, linux_net10.0, linux_net10.0_tr-TR, linux_net6.0, linux_net8.0, linux_net9.0, win32_net10.0, win32_net48, win32_net8.0, win_net10.0, win_net10.0_th-TH, win_net10.0_tr-TR, win_net462, win_net48, win_net48_de-DE, win_net48_fa-IR, win_net48_tr-TR, win_net5.0, win_net6.0, win_net6.0_ar-SA, win_net7.0, win_net8.0, win_net8.0_de-DE, win_net9.0, win_netcoreapp3.1, win_netcoreapp3.1_ja-JP: Chroma docs: "Forking is available in Chroma Cloud only"; a single server answers 501, which the client throws as a ChromaException

Schema: sparse vector search, BM25

• — a sparse vector index in the schema: Chroma docs: sparse vector search is in the Chroma Cloud section; a single server answers 400, which the client throws as a ChromaException
• ✓ a sparse vector index with BM25: written, read back, searched

Search API (basics, filtering, ranking, group by, hybrid, pagination, batch)

• — SearchAsync: Chroma docs: "Search API is available in Chroma Cloud only"; a single server answers 501, which the client throws as a ChromaException
• ✓ basics: knn, limit, select
• ✓ filtering: where, where_document, ids
• ✓ ranking: sums and products of rankings
• ✓ group by: the nearest of each category
• ✓ hybrid: RRF of a dense and a BM25 ranking
• ✓ pagination: limit and offset
• ✓ batch: several searches in one request

Quotas & Limits: the client keeps to them or reports the error of Chroma

• ✓ 300 records per write: the client splits; one request beyond gets the error
• ✓ 300 results: a get reads beyond in pages; a query or search beyond
• ✓ 10 reads or writes at a time per collection: 30 at once
• ✓ 8 predicates per filter
• ✓ 256 characters in a full-text search or a regex
• ✓ 4,096 dimensions
• ✓ 16,384 bytes per document
• ✓ 128 bytes per id, collection name and database name
• ✓ 8,182 bytes per metadata value
• ✓ 32 metadata keys

OpenTelemetry traces and metrics

• ✓ a span for each operation, with the attributes of the conventions
• ✓ the duration histogram and its buckets, with the OpenTelemetry SDK; — on win_netcoreapp3.1, win_netcoreapp3.1_ja-JP: the OpenTelemetry SDK exports no metric on this runtime, also from a meter of its own

Cancellation, timeouts, concurrent calls

• ✓ a canceled token, a timeout
• ✓ tasks at the same time on one client: 24, or 10 on Chroma Cloud

Dependency injection

• ✓ AddChromaClient, AddKeyedChromaClient, CreateChromaClient

Testcontainers module

• ✓ a Chroma container from ChromaDotNet.Testcontainers

Mocks in tests

• ✓ a subclass without a server

BM25 vectors computed by the client

• ✓ the vectors of the Python client of Chroma for 216 texts

Search and filter requests built by the client

• ✓ the JSON of ranks and filters