Chroma

Where Chroma sits

Chroma is the vector database most AI prototypes start with, and that is both its reputation and its business model. Founded in 2022 by Anton Troynikov and Jeff Huber in San Francisco, with an $18 million seed round in April 2023 led by Quiet Capital, it spread through the 2023 LLM boom on the strength of a single command — pip install chromadb — and deep integration with LangChain and LlamaIndex. The numbers attached to it are hard to argue with: over 26,000 GitHub stars, use in more than 90,000 open-source codebases, and over 11 million downloads a month, with production users including Capital One, Weights & Biases, UnitedHealthcare, Live Nation, Cisco, Intel and Sony.

Since August 2025 the story has a second half. Chroma Cloud, the managed serverless service built on a Rust engine over object storage, reached general availability, so the same codebase now runs three ways: embedded in your process, as a single server, or as a distributed cloud deployment. The open-source core stays Apache 2.0. In this category that combination — genuinely free self-hosting plus a usage-priced cloud with no minimum — is the reason Chroma keeps showing up in evaluations alongside Pinecone and Weaviate.

What it actually does

Chroma stores documents, their embeddings and their metadata in collections and retrieves them by vector similarity, by keyword, or by both. The query surface has grown well beyond "nearest neighbours": dense vector search for semantic similarity, sparse vector search with BM25 and SPLADE, full-text and trigram/regex matching, and metadata filtering with facets and GroupBy aggregation, all combinable as hybrid search. The core API is famously small — create a collection, add documents, query, get — and the system handles tokenisation, embedding and indexing automatically unless you bring your own vectors. Forking (copy-on-write collection duplication for dataset versioning and A/B tests) and read-level control over consistency are recent additions.

The ingestion story got aggressive in 2025-2026. Chroma Sync chunks, embeds and indexes GitHub repositories; Web Sync crawls, scrapes and embeds entire sites; there is a write-ahead log (wal3) built on object storage; and a Package Search MCP lets agents query thousands of open-source repos. The changelog is busy in the way that matters — Metadata Arrays, Indexing Status, private networking via AWS PrivateLink, customer-managed encryption keys — indicating a cloud product hardening toward enterprise use.

Chroma Cloud pricing is four honest meters with no feature gating: $2.50 per logical GiB written, $0.33 per GiB-month stored, $0.0075 per TiB queried, and $0.09 per GiB returned, plus small per-unit prices for Sync processing. New accounts get $5 in free credits. The Starter plan is $0/month plus usage (10 databases, 10 team members); Team is $250/month plus usage with $100 in included credits that do not roll over, Slack support, SOC 2 Type 2 and volume discounts; Enterprise is custom, adding single-tenant clusters, BYOC, dedicated support and SLAs. Self-hosting the Apache 2.0 core is completely free. The SOC 2 Type 2 certification covers AWS and GCP regions in the US and EU.

My testing notes

A few concrete observations.

First, the on-ramp claim is real. In my testing, the path from empty directory to a working vector store was an import statement and a create_collection call — no server, no Docker, no credentials — and the code written against the embedded client is the same code that runs against Chroma Cloud. That continuity is the strongest argument in the product; most competitors make you run something before your first document. For RAG experiments and agent-memory prototyping I have not found anything faster.

Second, the pricing is legible enough to model in a spreadsheet. The vendor's own calculator example — a million documents written, six million stored, ten million queries at 4 QPS — totals about $79/month. Query billing counts a vector similarity search as one query and each metadata or full-text predicate as additional queries (full-text and regex filters bill as N-2 queries for an N-character search string), so filter-heavy workloads need modelling before they need deploying. It is still a simpler bill than most rivals, but four meters is four meters.

Third, the single-node ceiling is physical and it shows up without warning. Capacity on one machine is bounded by memory — on the order of 245,000 records per gigabyte of RAM at 1024 dimensions — and query throughput parallelises only to the number of vCPUs before latency climbs with concurrency. In my use this never bit at prototype scale, and anyone who tells you it will not bite at production scale has not indexed enough documents. The distributed deployment behind Chroma Cloud is a different architecture with different latency and consistency behaviour, so what you measured locally does not predict cloud numbers.

The competitive shape is worth spelling out. Against Pinecone, Chroma trades maturity and multi-tenant isolation for a free local tier and a cloud bill with no minimum commitment. Against pgvector, it trades Postgres familiarity for purpose-built retrieval semantics and framework integrations that ship in tutorials by default. Against Weaviate and Qdrant, it trades feature depth and deployment control for the shortest path to a working demo. That last trade is the one most teams actually face in week one, and it is why Chroma keeps winning prototypes even when it loses final bake-offs. The practical consequence of that adoption is ecosystem gravity: LangChain, LlamaIndex and most agent toolkits treat Chroma as the default retriever, so the retrieval layer is rarely the part of a prototype anyone has to explain.

The embedded-to-cloud handoff deserves one warning paragraph, because it is the transition Chroma's whole strategy hinges on. The embedded library and the distributed service share an API and a licence, but they do not share performance characteristics, and the single-server mode is positioned for workloads under roughly ten million records across a handful of collections. Teams that validate locally and then move to Cloud are adopting a different consistency and latency model — plan the migration deliberately rather than discovering it under load.

What it does not do

The open-source server ships without built-in authentication or dependable multi-tenancy — a self-hosted deployment needs its own auth proxy and network controls before anything untrusted reaches it. Replication and high availability are not in the open-source offering; those arrive with Cloud and Enterprise. Multi-tenancy for true SaaS isolation is still catching up to Pinecone and Weaviate by the assessment of reviewers who tested it. Cold query latency on Cloud can reach roughly 1.5 seconds at p99 against warm-path queries, which matters for user-facing search. And the project has moved through major internal rewrites quickly enough that upgrades have historically involved data migrations and client changes — pin your version rather than floating it.

Two caveats on sources. One review claims Chroma lacks built-in hybrid search; that is out of date — sparse vectors (BM25/SPLADE), full-text and hybrid retrieval are documented first-class features since late 2025. And the vendor's "up to 10x cheaper" and cost-per-gigabyte comparisons are marketing claims, not measurements I could reproduce; I treat them as positioning.

Verdict

Chroma is the SQLite of the vector world, and I mean that as praise: it makes the first useful version of retrieval unusually easy and refuses to charge you for learning. I rate it 4.2 — weighted 40% core functionality depth (unified dense/sparse/full-text/metadata search, forking, sync pipelines, three deployment shapes), 20% value (free Apache 2.0 core, $5 cloud credits, no minimum, transparent meters), 20% usability (the fastest on-ramp in the category), 10% support and ecosystem (26k+ stars, 11M monthly downloads, framework ubiquity, active changelog), 10% reliability signals (Cloud GA only since August 2025, single-node ceilings, OSS auth gaps). For prototyping, local development and RAG pipelines up to serious size, it is my default recommendation; for multi-tenant SaaS isolation at massive scale, evaluate the dedicated managed players alongside it before committing.

Sources

• https://www.trychroma.com/pricing
• https://docs.trychroma.com/cloud/pricing
• https://github.com/chroma-core/chroma
• https://www.trychroma.com/
• https://www.aicentralresources.com/tool/chroma
• https://pecollective.com/tools/chroma/
• https://www.softwr.com/software/chroma
• https://www.itechguides.com/products/chroma/
• https://docs.trychroma.com/
• https://docs.trychroma.com/cloud