Pinecone · Weaviate · Qdrant · Milvus / Zilliz · Chroma · pgvector · Turbopuffer · MongoDB Atlas Vector · Vespa · LanceDB.
One question: which one is right for your stage?

1. Pinecone Series B+ · hosted leader · production-default · enterprise compliance posture

The hosted vector DB leader and production default — the substrate-of-choice when 'I want zero ops and I want it to work at scale' is the bar. Pinecone shipped serverless in 2024 (pay only for stored + queried vectors, no pod sizing) and remains the cleanest path from prototype-to-production for AI products that need real recall + real QPS without an ops team. Enterprise compliance posture is the strongest in the category: SOC 2 Type II, HIPAA BAA, GDPR DPA, AWS PrivateLink. The default hosted pick when 'two trillion-dollar companies wired by SideGuy' includes a memory substrate that doesn't break under production load. AI-baked-in (Pinecone was built for vectors from day one — never a general-purpose DB retrofitting vector search).

2. Weaviate Open-core · hybrid search leader · GraphQL · multi-tenant · cloud + self-host both

The open-core hybrid-search leader — best vector DB for teams that want GraphQL ergonomics, true hybrid (BM25 + vector) baked into the engine, and the option to self-host OR use Weaviate Cloud Services from one codebase. Weaviate's multi-tenant architecture (one cluster, thousands of isolated tenant namespaces) is the cleanest path for SaaS teams shipping AI features per-customer. Module ecosystem (text2vec-openai, text2vec-cohere, generative-anthropic) makes it the most opinionated 'batteries-included' OSS vector DB. AI-baked-in — built from day one for vector + hybrid search, not retrofitted.

3. Qdrant Series A · Rust-built · OSS-first · self-host favorite · cloud option

The Rust-built fast OSS vector DB with the cleanest self-host UX in the category — the operator's pick when 'I want to run this myself and have it not be painful' is the bar. Qdrant ships a single binary (Rust = no JVM, no Python runtime, no Docker hell) that runs on a laptop, a VPS, or Kubernetes with the same config. Strong filtering performance (vector search WITH metadata filters, not vector-then-filter), geo-search support, sparse vectors for hybrid. Qdrant Cloud exists for managed deployment but most operators choose Qdrant specifically because self-host doesn't suck. AI-baked-in — built specifically for vector workloads.

4. Milvus / Zilliz Zilliz · enterprise scale specialist · GPU-accelerated · billion-vector workloads

The enterprise-scale vector DB purpose-built for billion-vector workloads — GPU-accelerated indexing, distributed architecture, multiple index types (HNSW, IVF, DiskANN, GPU-CAGRA). Milvus is the OSS engine; Zilliz Cloud is the managed offering from the company that built Milvus. The right pick when 'we have 1B+ vectors, real QPS requirements, and we need the index to fit on disk because RAM cost is prohibitive' is the conversation. Zilliz Cloud Serverless launched 2024 for cost-control. AI-baked-in (purpose-built for vectors from day one).

5. Chroma Series A · embedded-first · dev-favorite · prototyping leader · cloud emerging

The dev-favorite embedded-first vector DB — the right pick when you want pip install chromadb, three lines of Python, and a working RAG prototype on your laptop in 5 minutes. Chroma's API is the simplest in the category (collection.add(), collection.query() — that's it). Runs embedded in your Python process, persists to local disk, no server to manage. Chroma Cloud launched 2024 for managed deployment, but Chroma's primary lane is prototyping, local-first AI apps, and embedded use cases (desktop apps, CLI tools, dev tools). AI-baked-in.

6. pgvector Postgres extension · OSS · zero-new-dependency · all-major-Postgres-providers

The Postgres extension that turns the database you already run into a vector DB — the right pick when 'one less dependency in our stack' beats 'best-in-class vector performance.' pgvector ships with Supabase, Neon, AWS RDS, Azure PostgreSQL, GCP Cloud SQL, and every major Postgres host. SQL syntax (SELECT ... ORDER BY embedding <-> query LIMIT 10), JOIN with your existing tables (no separate sync pipeline), transactional consistency between vector + relational data. HNSW index support since pgvector 0.5 brings recall-quality near purpose-built vector DBs at meaningful scale. AI-bolted-on architecturally (Postgres wasn't designed for vectors), but at <50M-vector scale the simplicity wins.

7. Turbopuffer Seed · serverless object-storage-backed · cheap-at-scale · cold-storage workloads

The serverless cheap-at-scale vector DB built on object storage (S3 / GCS) instead of always-on compute — the right pick when 'we have 100M-1B vectors but query rate is low and cost-per-vector matters more than sub-50ms latency' is the conversation. Turbopuffer indexes vectors on object storage with intelligent caching — pay for storage at S3 prices, pay for queries only when you query. Latency is higher than Pinecone (~100-300ms vs ~30-50ms) but cost-per-stored-vector can be 10-100x cheaper at large scale. AI-baked-in (built specifically for cheap vector workloads from day one).

8. MongoDB Atlas Vector MongoDB · Atlas integration · enterprise-default for MongoDB shops · 2024 GA

Vector search baked into MongoDB Atlas — the procurement-defensible pick when MongoDB is already the org default and adding a separate vector DB triggers a vendor review. MongoDB Atlas Vector Search ships HNSW index inside MongoDB Atlas clusters; same auth, same VPC peering, same compliance posture, same bill. JOIN-equivalent queries against your existing MongoDB documents (no separate sync pipeline). AI-bolted-on architecturally (MongoDB was never designed for vectors) but for MongoDB-native shops the procurement story dominates the technical tradeoff.

9. Vespa Yahoo-built · battle-tested at search-engine scale · OSS · hybrid lexical + vector

The Yahoo-built search engine that runs Yahoo Mail, Spotify recommendations, and other billion-document production workloads — vector search is one of many capabilities, not the whole product. Vespa is the most battle-tested engine in this list (it's been running production search at Yahoo scale for over a decade). True hybrid search (BM25 + vector + custom ranking signals + ML-ranking models) in one query. Best for teams that need lexical + vector + structured + ML-ranking fusion at billion-document scale. Operationally complex — not for solo founders. AI-bolted-on historically (added vector to a search engine), but the search-engine architecture is structurally suited.

10. LanceDB Seed · columnar Lance format · multi-modal · embedded · serverless emerging

The embedded multi-modal vector DB built on the columnar Lance file format — the right pick for multi-modal AI apps (image + text + audio + video) and embedded use cases that want object-storage economics. LanceDB stores vectors + metadata + raw multi-modal data in the Lance columnar format (designed for ML workloads — random access, versioning, SQL queries). Runs embedded in Python/JS/Rust like Chroma, but with much stronger multi-modal + analytics-grade querying. Lance format is the differentiator — same data accessible from PyArrow, DuckDB, Spark, Pandas without ETL. AI-baked-in.

🚀 If you're a Solo founder building an AI product (RAG over docs, semantic search)

Your problem: You're a solo or 2-3 person team shipping an AI product. RAG over your docs / customers' docs, semantic search over a product catalog, simple recommendation. Velocity matters more than 'best at scale.' You need a vector DB you can wire in 30 minutes and won't have to migrate off when you hit 1M vectors. Pair this decision with the AI Infrastructure megapage for the model substrate decision.

1. Pinecone — zero-ops hosted serverless — fastest path from prototype to production with no vector-DB ops to think about
2. Chroma — pip install chromadb + 3 lines of Python = working RAG prototype in 5 minutes; embedded, no server
3. pgvector — if you're already on Supabase / Neon / Postgres — one less dependency, JOIN with your existing tables
4. Qdrant — single Rust binary self-host that scales with you — cheap on a small VPS, painless to operate
5. LanceDB — if your AI product is multi-modal (image + text + audio) and you want columnar Lance format

📈 If you're a Series A startup adding AI features (production RAG · 1-10M vectors)

Your problem: You have product-market fit, paying customers, and you're adding AI features. 1-10M vectors, real QPS, sub-100ms latency budget, customer-data isolation matters. You need a vector DB that handles real production load AND has SOC 2 + privacy controls your enterprise customers will ask about. Pair with the Autonomous Coding Agents megapage for the build-velocity layer.

1. Pinecone — production-default hosted vector DB with strongest enterprise compliance posture (SOC 2 + HIPAA BAA + GDPR)
2. Weaviate — if you need true hybrid search (BM25 + vector) or per-customer multi-tenant isolation
3. Qdrant — self-host on Kubernetes if you can't send vectors to vendor cloud — single Rust binary, manageable ops
4. pgvector — if you're on Postgres and 1-10M is your ceiling — JOIN with relational data, transactional consistency
5. MongoDB Atlas Vector — if MongoDB is already the org standard — procurement-defensible, single Atlas bill

🏢 If you're a Mid-market integrating vector search into core product (50M+ vectors · hybrid filtering)

Your problem: You're 50-500 employees with 50M-500M vectors and real hybrid search needs (vector + keyword + metadata filters in one query). Your AI substrate has to clear a 4-12 week vendor onboarding process — SOC 2 Type II, DPA + data-residency + audit logs. Single-vendor lock-in is now a board-level risk. Coordinate with the Compliance Authority Graph for SOC 2 / ISO 27001 / HIPAA / GDPR posture.

1. Weaviate — true hybrid search (BM25 + vector fusion) baked into the engine + multi-tenant isolation + self-host or cloud both
2. Pinecone — hosted production-default with strongest enterprise compliance posture + serverless economics at this scale
3. Qdrant — self-host single Rust binary at this scale — strong filtered-vector-search performance, OSS inspectability
4. Milvus / Zilliz — if you're approaching billion-vector scale — GPU-accelerated indexing + multiple index types tuned per workload
5. Vespa — if hybrid lexical + vector + ML-ranking fusion at production scale is the core requirement

🏛 If you're a Enterprise CTO standardizing vector infra (multi-team · compliance · 100M+ vectors)

Your problem: You're 1000+ employees standardizing vector infrastructure org-wide. Multiple AI teams, multiple use cases (RAG, recommendation, semantic search, anomaly detection), multi-cloud reality (some teams on AWS, some on GCP, some on Azure). Strict procurement, central FinOps, audit + compliance + DPA + BAA. You're picking the substrate the next 5 years of AI products will be built on — AI-baked-in vs AI-bolted-on matters at this horizon (see /operator cockpit for the operator-layer view).

1. Pinecone — hosted enterprise default — strongest compliance posture, AWS PrivateLink, multi-region, proven at billion-vector scale
2. Milvus / Zilliz — Zilliz Enterprise + Zilliz Cloud — billion-vector GPU-accelerated, OSS Milvus on-prem option for regulated workloads
3. Weaviate — Weaviate Enterprise (cloud or self-host) — multi-tenant isolation across teams, hybrid search in engine, AI-baked-in
4. Vespa — if hybrid lexical + vector + ML-ranking at billion-document scale is a core team requirement (search-heavy orgs)
5. MongoDB Atlas Vector — rarely the enterprise vector standard — but defensible when MongoDB is already org-wide and procurement values bundle