The Numbers
The vector database market hit roughly $3.73B in 2026 with 23.5% annual growth per the consensus of MarketsandMarkets, Gartner, and IDC tracking. RAG is the primary demand driver. The “vector gold rush” of 2022-2024 settled into a clearer top tier — Chroma for prototyping and small workloads, Pinecone for managed enterprise reliability, Weaviate for hybrid search, Qdrant for performance, Milvus and Zilliz for the largest deployments, Faiss as the open-source library underneath many implementations, and pgvector for Postgres-native teams. The hyperscalers (AWS OpenSearch, Azure AI Search, GCP Vertex AI Vector Search) and Snowflake/Databricks native vector capabilities sit alongside as defaults for cloud-aligned shops.
Enterprise Leaders
Pinecone remains hard to beat on zero-ops enterprise reliability — the managed offering, the SOC 2 / HIPAA / FedRAMP coverage, the SLA. Qdrant wins on raw performance (typical p50 in the low-single-digit milliseconds for moderate corpora) and on cost at scale. Weaviate is strongest where hybrid search (vector plus BM25) is native rather than bolted on. Milvus through Zilliz Cloud handles the largest workloads and ships the strongest GPU-accelerated index types. pgvector continues to absorb low-traffic workloads in shops already running Postgres at meaningful scale. Pricing: serverless consumption replaced per-pod across most vendors in 2024-2025 — excellent for startup bursty workloads, risky for sustained high-query enterprise use without active cost monitoring.
Vendor selection — quick filter
Prototyping, small corpus Chroma, pgvector
Zero-ops enterprise SLA Pinecone
Hybrid search native Weaviate
Performance + cost at scale Qdrant
Largest workloads, GPU index Milvus / Zilliz
Salesforce-aligned Data Cloud native vector
Snowflake-aligned Snowflake Cortex SearchCRM Fit
For most CRM use cases, Salesforce Data Cloud’s native vector search, HubSpot’s Breeze knowledge base, or Microsoft’s Azure AI Search suffices. The vector tier is built into the CRM platform and the procurement, security, and ops are already in place. When you hit limits — tuning beyond the platform’s exposed knobs, cost at scale, multi-model embedding strategies, or specialized index types — evaluate Pinecone for simplicity, Qdrant for performance, pgvector if Postgres is already a core platform. The mistake to avoid: introducing a standalone vector DB before the platform’s native option has been proven inadequate against your actual workload.
Migration Path
Most CRM vector workloads do not hit the scale that forces a Milvus-tier choice. Start with native CRM search and migrate only when measurable: latency exceeds the SLA at peak, cost-per-query exceeds the budget envelope, recall on the eval set is below threshold, or multi-model embedding requirements outgrow the native tier. Premature optimization on vector DB choice costs more than it saves and locks engineering capacity into platform work that the team did not need to do.
What Changed in 2026
Three shifts: serverless consumption pricing became the dominant model, requiring active cost management; embedding models (OpenAI text-embedding-3-large, Cohere Embed v4, Voyage AI) accelerated, raising the cadence of corpus re-embedding; and the platform-native vector tier (Salesforce Data Cloud, HubSpot, Microsoft Fabric) matured to the point that for many CRM use cases the standalone vendor is no longer the default first choice.
Common Failure Modes
The recurring failures: starting on serverless pricing for a known-high-volume workload because “we will optimize later”; missing the budget alert because nobody owns the line item; over-fetching top-k=20 when top-k=5 would suffice; running multiple overlapping retrievers per turn without measuring incremental value; and locking on a single vector DB before measuring whether the platform-native option suffices.
Cost Considerations
Pinecone Serverless pricing in 2026 sits around $0.33 per million read units; Qdrant Cloud and Weaviate Cloud comparable. Storage roughly $0.33/GB/month. Re-embedding the corpus on a model upgrade is a one-time spike worth budgeting separately. Plan a quarterly cost review with the AI platform team.
What to do this week
Pull last month’s vector-spend invoice and divide by the conversation count. The cost-per-conversation number is the unit-economics input every architectural decision should defer to.
