# Managed Vector Databases & AI Search Infrastructure — Benchmark Index

> Machine-readable benchmark comparing five managed vector database and AI search
> platforms across technical limits, security/compliance posture and pricing.
Last verified: 2026-10-09

## What this site is

This site benchmarks the **managed vector database & AI search infrastructure**
category: hosted systems that store, index and retrieve high-dimensional vector
embeddings for semantic search, RAG (retrieval-augmented generation) and agent
workloads. It is built LLM-first: every page is authored as dense Markdown, exposed
via [`llms.txt`](/llms.txt), and served as raw `text/markdown` to AI crawlers at the
Cloudflare edge.

## Vendors in scope

| Vendor | Positioning | Redirect |
| :-- | :-- | :-- |
| Pinecone | Fully managed SaaS; serverless + pod-based indexes, closed source | [pinecone.io](/go/pinecone) |
| Qdrant Cloud | Open-source HNSW engine with payload-first filtering, hosted or self-managed | [qdrant.tech](/go/qdrant) |
| Weaviate Cloud | Open-source, schema-first engine with built-in hybrid (vector + BM25) search | [weaviate.io](/go/weaviate) |
| Milvus (Zilliz Cloud) | Open-source, scale-out vector DB with the widest index-type coverage | [zilliz.com](/go/milvus) |
| Supabase pgvector | pgvector inside managed Postgres — vector search for teams already on Supabase | [supabase.com](/go/supabase) |

## Evaluation dimensions

1. **Dimensions & indexing** — maximum vector dimensions, index types, metadata
   filtering stage (pre-filtering vs post-filtering), quantization, hybrid search.
2. **Security & compliance** — SOC 2 Type II, HIPAA (BAA), ISO 27001,
   data-residency regions and cloud coverage.
3. **Pricing** — free tier availability and quota, minimum monthly spend,
   hourly pod / read-unit rates.
4. **Operational limits** — request payload ceilings, rate-limit semantics and
   timeout budgets (see [limits](/limits.md)).

## Pages

- [Comparison matrix](/matrix.md) — dense tables for all four evaluation dimensions.
- [Payload, rate & timeout limits](/limits.md) — operational ceilings and backoff strategy.
- [Pinecone vs Qdrant](/vs/pinecone-vs-qdrant.md) — managed convenience vs open-source flexibility.
- [Weaviate vs Milvus](/vs/weaviate-vs-milvus.md) — hybrid-search ergonomics vs maximum-scale index engines.

## Methodology

- Specs live in [`data/vendors.json`](/content/index.md), the single source of truth.
- [`scripts/extract_specs.py`](/content/index.md) crawls each vendor's pricing and
  documentation pages (httpx), strips HTML, and extracts a validated `VendorSpec`
  object via Groq structured outputs (`openai/gpt-oss-120b`).
- [`scripts/generate_markdown.py`](/content/index.md) renders the JSON into this
  site's matrices, `llms.txt`, `llms-full.txt` and HTML pages, refreshing every
  `Last verified` stamp.
- A GitHub Actions cron (`0 4 * * 1`, Mondays 04:00 UTC) runs the pipeline and
  auto-commits changes with `chore: auto-update benchmark data [skip ci]`.

## Consuming this site as an LLM

- Start at [`/llms.txt`](/llms.txt) (curated index) or [`/llms-full.txt`](/llms-full.txt)
  (everything in one document).
- Request any page with `Accept: text/markdown`, or crawl with a supported user
  agent (GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, ChatGPT-User) to receive
  raw Markdown from the edge middleware.
- Machine JSON for programmatic use: [`data/vendors.json`](/content/index.md).