Chroma

Chroma is an open-source vector database developed by Chroma Inc. and first released in October 2022. It stores embeddings, the documents they were derived from, and arbitrary metadata, and provides similarity search APIs designed for retrieval-augmented generation, semantic search, recommendations, and other embedding-driven applications. Distributed under the Apache 2.0 licence, Chroma is widely used in prototypes built with LangChain and LlamaIndex and has, since a 2025 Rust-core rewrite, become viable for larger production deployments.

Architecture

Chroma exposes a Python and JavaScript client API organised around collections, which group embeddings, source documents, and metadata. Each collection is associated with an embedding function — Chroma can either compute embeddings itself using built-in integrations (OpenAI, Cohere, sentence-transformers, Hugging Face, Voyage, Jina, and others) or accept pre-computed vectors. Queries return the nearest neighbours by cosine, L2, or inner-product distance, with optional metadata filtering using a familiar where-clause syntax. Earlier versions of Chroma ran as an in-process Python library backed by SQLite and DuckDB, optimised for local development and small workloads. The 2025 Rust rewrite kept that single-file ergonomics but added a true multithreaded core that the company reports as roughly four times faster on writes and queries than the previous Python implementation, removing the Global Interpreter Lock as a bottleneck.

Chroma Cloud and serverless

Chroma Cloud is a managed serverless deployment introduced in 2025 with an architecture that separates query nodes from compactor nodes, using cloud object storage as the shared persistence layer. Query nodes serve indices directly from object storage while compactor nodes build and update them asynchronously, which lowers operational cost compared with SSD-replicated clusters at the expense of slightly higher cold-start latency. Customer data can be protected with customer-managed encryption keys and AWS PrivateLink connectivity.

Search features

Chroma supports dense vector search, BM25 lexical search, and SPLADE sparse vector search, allowing applications to combine semantic and keyword retrieval in a single query. Recent additions include regular-expression search over document text, a GroupBy operator that returns top results per metadata bucket, and richer metadata types including arrays of strings, numbers, and booleans. The simple collection model and zero-configuration local mode make Chroma a popular default for developers prototyping retrieval-augmented generation pipelines.

Comparison with alternatives

| Database | First release | Licence | Strength | | --- | --- | --- | --- | | Chroma | 2022 | Apache 2.0 | Developer experience, embedded mode | | Pinecone | 2021 | Commercial SaaS | Fully managed, large index scale | | Weaviate | 2019 | BSD-3 | Hybrid search, GraphQL API | | Qdrant | 2021 | Apache 2.0 | Quantisation, advanced filtering | Chroma is typically the right choice for rapid development and workloads up to roughly ten million vectors, while Pinecone, Weaviate, and Qdrant are usually preferred for very large indexes, advanced security and access control, or specialised quantisation needs.

Ecosystem integration

Chroma is a first-class retriever in LangChain, LlamaIndex, Haystack, and many agent frameworks, and is included as a default option in the OpenAI and Anthropic cookbooks for retrieval-augmented generation. It is also a common backing store for local AI assistants packaged with desktop applications because the embedded mode requires no separate server process. Chroma is one of the most commonly adopted vector databases among Malaysian developers and startups building retrieval-augmented generation systems on local content, including Bahasa Malaysia knowledge bases, regulatory libraries, and customer-support corpora. Its embedded mode runs comfortably on developer laptops and small VMs, which suits the tight cloud budgets typical of early-stage Malaysian AI products. Local fintechs and SaaS firms use Chroma alongside Amazon Bedrock and Anthropic Claude models hosted via VSTECS-billed AWS accounts to build retrieval-augmented chat over policy documents, product catalogues, and Bursa Malaysia filings. The simple Python and JavaScript API also fits the workflow of MMU, UM, and APU students prototyping AI projects under MDEC's #mydigitalmaker programme. For regulated workloads in banking, insurance, or healthcare, Malaysian system integrators typically pair Chroma Cloud's customer-managed key and PrivateLink features with the Bank Negara Malaysia RMiT and Personal Data Protection Act (PDPA) 2010 requirements on encryption and data residency. Where stricter on-premises hosting is mandatory, MIMOS Berhad reference architectures show Chroma running on Kubernetes inside Cyberjaya colocation facilities. AITG Sdn Bhd's Teragrid Agent and Teragrid Ai Platform list Chroma among the supported vector backends for partner-deployed retrieval-augmented generation, alongside Pinecone, Qdrant, and Weaviate.

1. Chroma Inc. (2025). Chroma documentation. trychroma.com.
2. Adyog Blog (2025). Chroma Vector Database: The Open-Source Foundation for AI Search.
3. Firecrawl (2026). Best Vector Databases in 2026: A Complete Comparison Guide. firecrawl.dev.
4. Airbyte (2025). Chroma DB vs Qdrant: Key Differences. airbyte.com.