AI Governance & Security Vendor Wiki

9 min read

Vector databases: the landscape

A vector database stores embeddings and answers “what is semantically near this query” — the index layer between your content-sources and whatever assistant or agent consumes the retrieved snippets. The wiki files this under vector-retrieval, a deliberately thin, situational category with no survey question: which vector DB you pick is an engineering decision, not a governance control (decision D9/Q3 in taxonomy-gaps.md). The control that does matter sits one layer up, in entitlement-aware-rag.

This page is the landscape scan behind that stance — filed back from a scoping query (2026-08-24), covering the vendors that come up in every RAG conversation but do not each earn a vendor page. Four already have pages: pinecone, weaviate, azure-ai-search, opensearch.

The market has four shapes

Almost every option is one of these, and the shape matters more than the benchmark:

  1. Managed pure-play SaaS — Pinecone, Weaviate Cloud, Zilliz Cloud, Qdrant Cloud. You send embeddings to someone else’s cloud.
  2. Self-hosted OSS engines — Qdrant, Milvus, Weaviate OSS. You run a stateful distributed database; embeddings never leave your tenancy.
  3. Embedded / in-process libraries — Chroma, LanceDB. The index is a file or a library inside your app, not a server you operate.
  4. “Use what you already have” — pgvector on Postgres, OpenSearch/Elasticsearch, MongoDB Atlas Vector Search, Redis, Azure AI Search, S3 Vectors. No new vendor, no new data store, no new DPA.

For a fund with an existing Postgres or Elastic estate, shape 4 is usually the right default and shapes 1–3 need a reason.

The pure-plays

As of 2026-08. Funding and ownership figures are dated because they move.

ProductShapeLicenseOwnership / funding (as of 2026-08)Notes
pineconeManaged SaaS onlyProprietaryIndependent; ~$138M raised, $100M Series B (2023) at ~$750M. Reportedly exploring a sale — see M&A belowThe original managed pure-play; serverless, lowest ops burden; not self-hostable, so embeddings live in their cloud
weaviateOSS + managed cloudBSD-3Independent; Series B $50M (2023), Series C (2025); RICOH Innovation Fund investment 2026-03-13Hybrid (vector + keyword) search, GraphQL API; self-host is the residency story
QdrantOSS + Cloud / Hybrid Cloud / EdgeApache 2.0Independent; $50M Series B announced 2026-03-12 (led by AVP; Bosch Ventures, Unusual, Spark, 42CAP), ~$87.8M totalRust engine; the performance/filtering pick. Hybrid Cloud runs the data plane in your infrastructure with their control plane — the interesting shape for regulated buyers
Milvus / ZillizOSS + Zilliz CloudApache 2.0 (LF AI & Data project)Zilliz ~$113M raised; last round a $60M Series B extension (2022-08); ~139 employees (2026-03)The billion-scale option; heaviest to operate. Vendor funding is old — worth watching
ChromaEmbedded-first + Chroma CloudApache 2.0Independent; $18M seed (2023, led by Quiet Capital) — no later round foundThe developer-experience/prototyping default; pip install and you have an index. Chroma Cloud (serverless on object storage, Rust core) GA 2025-08
LanceDBEmbedded + cloud, on the Lance columnar formatOSS coreIndependent; $30M Series A (2025-06)Multimodal (“lakehouse”) angle — text, image, video, plus SQL and full-text over the same files; used as an agent-memory substrate (e.g. CrewAI), Netflix and ByteDance cited as users

The substrate options

Not pure-plays; usually the cheaper and more governable answer.

OptionWhy it shows up
pgvector (Postgres)The 2026 default for production RAG at fund scale. Vectors sit next to relational data, inherit your backup/DR/audit posture, and Postgres row-level security can enforce tenant/entitlement isolation inside the query (see the pre- vs post-filter trap below)
opensearch / ElasticsearchHybrid keyword+vector in an engine most shops already run for logs
azure-ai-searchAzure-native; supports security-trimming via document-level ACLs — the closest thing to entitlement-aware retrieval you get for free
MongoDB Atlas Vector Search, RedisSame “already in the estate” logic
Amazon S3 VectorsNative vector storage in S3; GA 2025-12, expanded to 17 more regions 2026-03. AWS claims up to 90% cost reduction (vendor figure). Aimed at cold/large corpora, not low-latency serving
turbopufferObject-storage-native search engine; vendor cites 2.5T+ vectors and ~$0.02/GB storage. Same architectural bet as S3 Vectors and Chroma Distributed
Vespa, Vertex AI Vector SearchLarge-scale hybrid search; GCP-native equivalent of Azure AI Search

One layer up: RAG platforms (not vector DBs)

These get named alongside vector databases but are a different purchase — they bundle retrieval, generation, and (increasingly) governance. They belong nearer enterprise-ai-assistant and entitlement-aware-rag than to this page:

  • Vectara — managed RAG/agent platform whose differentiator is hallucination detection: the Hughes Hallucination Evaluation Model (HHEM) and its public leaderboard, with HHEM-2.3 behind the API and HHEM-2.1-Open weights published. ~$53.5M raised ($25M Series A led by FPV Ventures and Race Capital). The measurable-groundedness angle also touches llm-observability.
  • Contextual AI — “RAG 2.0” enterprise RAG-agent platform from Douwe Kiela and Amanpreet Singh (ex-FAIR/Hugging Face; Kiela led the team behind the original 2020 RAG paper). ~$100M raised ($80M Series A); sells into banking and financial services (HSBC, Qualcomm cited).
  • Glean already has a page under entitlement-aware-rag — the permissions-mirroring index is the governance-relevant version of this same idea.

Neither Vectara nor Contextual AI has a wiki page yet; both are plausible candidates and are flagged in open questions.

Consolidation / M&A dynamics

The standalone vector DB is under pressure from both directions — bundling from below (every OLTP and search engine added vectors) and object-storage economics from above.

  • Pinecone is reportedly exploring a sale. Secondary coverage (via The Information) says bankers were engaged, with speculation of a valuation above $2B against its last $750M mark, and Oracle, IBM, MongoDB and Snowflake floated as buyers. Founder Edo Liberty moved from CEO to Chief Scientist on 2025-09-08, with Ash Ashutosh (founder of Actifio, acquired by Google 2020) named CEO; Ashutosh has publicly said an acquisition is not the goal. Unconfirmed — no transaction announced as of 2026-08-24.

    Note (soft): this is press reporting, not a primary source, and “exploring a sale” is not a deal. Tracked as a watch item, not an M&A state change.

  • The bear case is now said out loud. Elastic CEO Ashutosh Kulkarni: “Vector databases are a feature. They are never going to be a business in and of themselves.” Self-interested, but it matches the pgvector-wins pattern.

  • Money is still flowing to the OSS engines — Qdrant’s $50M Series B (2026-03), Ricoh’s strategic investment in Weaviate (2026-03) — which suggests the survivors are the ones with a self-host story, not the managed pure-plays.

  • Zilliz’s last disclosed round is 2022. Not a red flag by itself for an OSS project with a foundation home, but worth noting before making Milvus a dependency.

For a buyer, the practical read: prefer options with a credible exit path. Apache-2.0 engines and Postgres extensions survive their vendors; a proprietary managed API does not.

What actually matters for governance

The database choice is mostly not a security decision. Four things are:

  1. The index is a copy of your corpus with the ACLs stripped off. Permissions lived on the SharePoint folder, not on the embedding. This is the classic RAG failure mode and it is a property of the pipeline, not the DB. See entitlement-aware-rag-options.
  2. Pre-filter, not post-filter. Enforcement must happen inside the similarity query (namespace, metadata filter, Postgres RLS) — filtering the top-k after retrieval both leaks timing/existence signal and produces the “empty results” failure that tempts engineers to loosen the filter.
  3. Embeddings are sensitive data. They are a lossy but partially invertible representation of the source text, so an embedding store holding MNPI is an MNPI store: it needs the same residency, encryption/BYOK, retention, and deletion treatment as the documents. This is why the shape (managed SaaS vs self-hosted vs embedded) is the decision that matters — it determines whether sensitive content leaves your tenancy at all. Adjacent: dspm.
  4. Deletion has to propagate. Removing a document from the system of record does nothing to its embedding, its chunks, or any agent memory built on top of it — see agent-memory-landscape, where the same store reappears as the persistence layer for agent recall.

Choosing, in one paragraph

If you already run Postgres, start with pgvector and stop; if you already run Elastic/OpenSearch, start there. Reach for a pure-play only on a specific trigger: scale past what Postgres serves comfortably, or a genuine need for hybrid/filtered search performance (Qdrant), or billion-vector scale (Milvus), or multimodal corpora (LanceDB). Choose Pinecone for lowest operational burden if you accept third-party hosting of embeddings — a real MNPI/residency question for a fund, and now also a vendor-continuity question. Use Chroma for prototypes and expect to migrate.

Open questions

  • Do Vectara and Contextual AI earn vendor pages? Both make governance-relevant claims (measurable groundedness; regulated-industry RAG). Neither has been verified against primary sources. Probable homes: enterprise-ai-assistant / entitlement-aware-rag, with llm-observability cross-listing for Vectara.
  • Should Qdrant, Chroma, LanceDB, Milvus get stub vendor pages, or does this page discharge the coverage obligation? Current answer: this page — they add nothing to the governance story that weaviate does not already carry.
  • pinecone needs a watch date on the sale reporting; page updated 2026-08-24 with verify_after.
  • Funding/pricing figures here are from secondary sources and vendor blogs, not cached primary filings. Promote and cache before any claim moves onto a vendor page.

Sources

Accessed 2026-08-24; secondary/roundup-heavy except where noted. Two cached to raw/sources/ (the Pinecone and Qdrant items, which back page edits):

Graph View

Backlinks