pgvector vs Milvus
Five of fourteen models named pgvector first on the direct prompt; zero named Milvus. Both were named by all fourteen models and pgvector carries 73 labels and Milvus 43, so the shares are not directly comparable.
Named in one category this edition.
Named in one category this edition.
Share is the count of first choices across the direct, paraphrase, budget and scale prompts over all fourteen models, for a mid-market B2B company; rank is within the category; every quote names the model and the prompt it came from. Both figures come from the vector databases page.
By framing
pgvectorFirst choices, of fourteen modelsMilvus
Across every category in the November 2026 Edition, pgvector and Milvus were named in the same answer eighty-eight times, of the 192 answers naming pgvector and the 114 naming Milvus. In those answers Milvus took the first choice four times and pgvector twenty-four.
Every model, every framing
| Model | Direct | Paraphrase | Comparative | Budget-constrained | Scale-constrained | Negative |
|---|---|---|---|---|---|---|
| Claude Haiku 4.5 | | | | | ||
| GPT-5.4 mini | | | | | ||
| Gemini 3.5 Flash | | | | | | |
| Perplexity Sonar | | | | | | |
| Grok 4.20 | | | | | | |
| Mistral Small | | | ||||
| DeepSeek V4 Flash | | | | | | |
| Llama 4 Maverick | | | | | ||
| Qwen 3.7 Flash | | | | | | |
| Kimi K2 | | | | | | |
| GLM 4.7 FlashX | | | | | ||
| MiniMax M2.5 | | | | | | |
| GPT-6 Luna | | | | | | |
| Muse Glimmer 30B | | | | | | |
pgvector Milvus first choice named as an alternative argued againstblank: not namedEach cell is one answer, pgvector on the left and Milvus on the right.The direct prompt
pgvector first, Milvus an alternative
5 of 14 modelsMilvus was named in the answer but not as the choice, or not at all.Neither was the first choice, one was named
9 of 14 modelsThe answer put something else first and named one of the two as an alternative.Bold names in an answer are the products the judge labeled a first choice; a model naming several gives each of them that label. The full answer text for every row is in the record.
By buyer segment
pgvector35%#1 of 9Milvus1%#6 of 9pgvector34%#2 of 11Milvus2%#5 of 11pgvector7%#4 of 13Milvus7%#5 of 13What the models said about pgvector
Every negative label with a quote, up to three, then the highest-weighted positives, up to three. Six of eight in this category shown.
“pgvector does not natively support GPU acceleration or disk-optimized indexes, and at billion-scale workloads typically require more advanced sharding strategies” Claude Haiku 4.5 · negative prompt · soft negative
“Using vectors beside your application data in Postgres via pgvector is often the safest default until measured retrieval requirements force a dedicated engine.” Muse Glimmer 30B · negative prompt · soft negative
“pgvector for large, high-throughput workloads with selective filters or many tenants... approximate-index filtering can return too few results” GPT-6 Luna · negative prompt · soft negative
“the recommendation is to start with pgvector unless your scale, query patterns, or operational constraints provide specific evidence otherwise” Claude Haiku 4.5 · scale prompt · first choice
“Start simple: Use pgvector (in Postgres) or SQLite with vector extensions for most apps—it's battle-tested, avoids new tech debt” Grok 4.20 · negative prompt · first choice
“If your company already uses PostgreSQL for standard application data, pgvector is almost always the most cost-effective choice.” Gemini 3.5 Flash · budget prompt · first choice
What the models said about Milvus
Every negative label with a quote, up to three, then the highest-weighted positives, up to three. Six of eight in this category shown.
“Milvus for small or moderate workloads: ... it can require substantially more memory and operational complexity than lighter alternatives” Perplexity Sonar · negative prompt · soft negative
“If you need a more scalable, distributed solution later, but it's overkill for early stages.” GLM 4.7 FlashX · budget prompt · soft negative
“Handles massive scale (10B+ vectors) but higher memory use (up to 2.6GB in benchmarks)” Grok 4.20 · budget prompt · soft negative
“If you have a platform/DevOps team: Evaluate Milvus or Qdrant self-hosted for cost control at scale” Kimi K2 · scale prompt · first choice
“Open-source, battle-tested at scale (used by big tech), and offers managed cloud (Zilliz Cloud).” MiniMax M2.5 · direct prompt · alternative
“Expect a large distributed deployment or need extensive index choices? Evaluate Milvus.” GPT-6 Luna · comparative prompt · alternative
Comparisons are drawn for the top eight products in each category, each against each. The output is the models' output; nothing here is a recommendation by the index.