Bottom line
Inference, medium confidence: adopt and wrap an existing vector store rather than writing an index. Use pgvector if you already run Postgres, an embedded library (Chroma) for single-process apps, or a Qdrant server at scale, with faiss as the exact-search reference; Milvus, Weaviate, LanceDB and Vespa are watch-level options whose CI and tests were not read. Wrap it with your own recall check against exact search on your data, including your metadata filters, before trusting any number: the public benchmark harness most server vendors submitted to is no longer maintained, and no dedicated filtered-recall suite was verified in any vendor repo (only Qdrant's filter-relevant tests were checked).
Adopt, do not rebuild
License checked on 2026-09-23; non-permissive licenses are named in the text.
- facebookresearch/faiss MIT: the ANN (approximate nearest neighbour) library the companion names as the exact-search ground-truth oracle, and the reference for your own recall checks. A fresh read on 2026-09-23 of https://api.github.com/repos/facebookresearch/faiss/contents/.github/workflows listed
build-pull-request.yml,nightly.ymlandindex-io-backward-compatibility.yml. [Verified] (ecosystem/pickup/pickup-vector-dbs.md:76reference implementation / ground-truth
; ecosystem/pickup/_evidence/vector-dbs.md:9Canonical ANN library everything else is measured against
) - pgvector/pgvector PostgreSQL: HNSW and IVF indexes inside Postgres, for teams that would rather extend their database than deploy a new one. A fresh read on 2026-09-23 of https://raw.githubusercontent.com/pgvector/pgvector/master/.github/workflows/build.yml showed a build matrix across Postgres 13 to 20. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:12
Vector ANN inside Postgres (HNSW/IVF)
) - qdrant/qdrant Apache-2.0: Rust HNSW server that submits parameter-sweep configs to the public ANN benchmark and runs lint, coverage and long end-to-end CI. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:10
submits parameter-sweep configs to ANN-Benchmarks
; ecosystem/pickup/_evidence/vector-dbs.md:34Rust-specific CI: lint + fmt + clippy + coverage + nightly E2E
) - milvus-io/milvus Apache-2.0: cloud-native vector database with client conformance suites for several languages under one test tree; Watch-level: existence, activity and file layout only, its CI and tests were not read. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:8
cloud-native vector database
; ecosystem/pickup/_evidence/vector-dbs.md:32Multi-language client conformance suites under one tests/ tree
) - chroma-core/chroma Apache-2.0: embedded-first store with property-based correctness tests. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:11
embedded-first alternative
) - weaviate/weaviate BSD-3-Clause AND LicenseRef-Weaviate, lancedb/lancedb Apache-2.0 and vespa-engine/vespa Apache-2.0: a Go server with hybrid BM25 search, an embedded multimodal library with a Rust core, and a search platform that unifies ANN with lexical ranking. All three are Watch-level: existence, activity and file layout only; their CI and tests were not read. Weaviate is open core: code outside
wl/is BSD-3-Clause, whilewl/is under the proprietary Weaviate License ('All rights reserved') and runs only with aLICENSE_KEY; without a key only BSD-3-Clause code executes. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:13GO vector DB with modules (hybrid search/BM25)
; ecosystem/pickup/_evidence/vector-dbs.md:14Embedded multimodal retrieval library
; ecosystem/pickup/_evidence/vector-dbs.md:15AI search platform with ANN + lexical ranking unified
; stack/licenses.tsv:100open core, dual-licensed
; stack/licenses.tsv:100wl features run only with LICENSE_KEY
; stack/licenses.tsv:100only BSD-3-Clause-licensed code is executed
)
Copy these practices
- Measure recall against exact search, on your data: use faiss exact search to generate ground truth, then report recall at k for the configuration you run. Starter kit: Incumbent oracle pinned before implementation. [Inference] (ecosystem/pickup/pickup-vector-dbs.md:77
exact-search ground truth generation
) - Report a recall-versus-speed curve from a parameter sweep, not one number: the ann-benchmarks
algorithms/qdrant/config.ymlgrid over graph degree, build and query effort, quantization and re-scoring. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:25Recall-vs-QPS Pareto via explicit parameter sweeps
) - Test filtered queries separately: filtered recall needs its own query set with precomputed answers (the companion's proposed GATE-VDB-02). Starter kit: none. [Inference] (ecosystem/pickup/pickup-vector-dbs.md:163
filtered ANN claims require a dedicated
) - Property-based tests with written invariants: Chroma
chromadb/test/property/; a fresh read on 2026-09-23 of https://raw.githubusercontent.com/chroma-core/chroma/main/chromadb/test/property/invariants.py showed invariants such ascount,ids_match,embeddings_match,no_duplicatesand an_exact_distanceshelper. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:27Property-based correctness testing (Hypothesis strategies + explicit invariants)
) - Stress tests for low memory and low disk, apart from unit tests: Qdrant
tests/e2e_tests/test_low_ram.py,test_low_disk.py. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:28Stress / soak suites separate from unit tests
) - Fail CI when generated API specs drift from their sources: Qdrant
tests/openapi_consistency_check.sh(fresh read on 2026-09-23 of https://raw.githubusercontent.com/qdrant/qdrant/master/tests/openapi_consistency_check.sh: its header says it makes sure generated OpenAPI files are consistent with their sources). Starter kit: none. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:29API-contract regression (OpenAPI consistency checks)
) - Ship the benchmark tool in the main repo: Vespa's
vbench/. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:31Ships a perf-regression CLI inside the main repo
) - Write contributor instructions for coding agents: LanceDB, Milvus, Weaviate and Chroma keep AGENTS.md or CLAUDE.md files. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/vector-dbs.md:37
AGENTS.md / CLAUDE.md contributor-playbook docs
)
Build only if
- Your index must record which embedding model produced its vectors and refuse to mix embedding spaces, in-process with no server. frankensearch's index does this; its packet judges it more provenance discipline than most vector databases ship, which is an analyst judgment, not a survey. [Inference] (packets/frankensearch-assessment.md:100
Identity checks fail closed if vectors from the two spaces are combined
; packets/frankensearch-assessment.md:168more provenance discipline than most vector DBs ship
) - License is not a reason to build an index: faiss, pgvector, Qdrant, Milvus, Chroma, LanceDB and Vespa are MIT, PostgreSQL or Apache-2.0, and Weaviate's proprietary terms cover only its key-gated
wl/features, so they justify choosing another store, not writing one. [Verified] [Inference] (stack/licenses.tsv:23facebookresearch/faiss MIT permissive
; stack/licenses.tsv:74pgvector/pgvector PostgreSQL permissive
; stack/licenses.tsv:80qdrant/qdrant Apache-2.0 permissive
; stack/licenses.tsv:13chroma-core/chroma Apache-2.0 permissive
; stack/licenses.tsv:60milvus-io/milvus Apache-2.0 permissive
; stack/licenses.tsv:43lancedb/lancedb Apache-2.0 permissive
; stack/licenses.tsv:98vespa-engine/vespa Apache-2.0 permissive
; stack/licenses.tsv:100wl features run only with LICENSE_KEY
)
Where FrankenSuite touches this
- frankensearch-index: an f16 SIMD vector index with brute-force top-k, memory-mapped reads and frozen producer fingerprints, published separately on crates.io; part of frankensearch, TRL 5-6, NODUS ring Explore. It requires the pinned nightly toolchain, which excludes stable-Rust consumers, and carries the license rider. [Verified] [Inference] (packets/frankensearch-assessment.md:289
the pinned nightly-2026-08-31 toolchain (stable-Rust consumers are excluded)
; packets/frankensearch-assessment.md:92f16 SIMD vector index (FSVI format
; packets/frankensearch-assessment.md:289independently published on crates.io as 0.3.1 with 4,148 downloads
; packets/frankensearch-assessment.md:22TRL 5–6 — see §4.9). Substantive and partially validated
) - Approximate search there is held in reserve: HNSW through a pinned fork is feature-gated, and the latency receipts cover only brute force. Both vector generations carry RaptorQ error-correction sidecars, whose recovery was not executed by the assessor. [Verified] (packets/frankensearch-assessment.md:291
the receipts only cover brute-force top-k
; packets/frankensearch-assessment.md:99RaptorQ FEC sidecars for both vector generations
) - The packet names Qdrant, Weaviate and Milvus as the incumbents it does not contest, and its license rider bars OpenAI, Anthropic and anyone acting for them. [Inference] [Verified] (packets/frankensearch-assessment.md:148
with filtering, quantization, and distributed stories frankensearch doesn't contest
; packets/frankensearch-assessment.md:175The license bars OpenAI, Anthropic, affiliates, and anyone acting for them
) - frankengraphdb designs a hybrid ANN, BM25 and graph retrieval operator, which is not integrated. [Maintainer claim] (packets/frankengraphdb-assessment.md:117
the hybrid ANN+BM25+graph operator is not integrated
)
What we cannot say
- Whether any vendor's filtered search returns the right results: only filter-relevant test files were found (ecosystem/pickup/_evidence/vector-dbs.md:43
Filter-correctness file evidence is partial
). - How the systems compare now: the public harness is sunset and its successor is small (ecosystem/pickup/_evidence/vector-dbs.md:41
and points to vector-index-bench/vibe (73 stars, early)
; ecosystem/pickup/pickup-vector-dbs.md:347Is VIBE (73 stars, early) stable enough
). Chroma and LanceDB never submitted to it, so their numbers are self-hosted (ecosystem/pickup/_evidence/vector-dbs.md:44Chroma and LanceDB are NOT in ann_benchmarks/algorithms/
). - Who runs these in production: descriptions come from repo metadata (ecosystem/pickup/_evidence/vector-dbs.md:45
I did not verify named company adopters, so none are claimed
). - How deep LanceDB's Python tests go (ecosystem/pickup/_evidence/vector-dbs.md:46
lancedb python/tests/ currently only holds test_oauth.py + test_otel.py
), or what Vespa'svbenchcovers, since its README path 404'd (ecosystem/pickup/_evidence/vector-dbs.md:47vespa vbench/README 404'd on master
). - How to verify hybrid dense-plus-keyword results: no independent ground truth was identified (ecosystem/pickup/pickup-vector-dbs.md:368
Ground-truth provenance for hybrid (dense+sparse) tracks
).
Revisit when
- VIBE tags a release with participant modules from the major vendors.
- A dedicated filtered-recall test suite with precomputed ground truth is verified in any vendor repo.
- frankensearch receipts its HNSW path.
Sources
- Pickup companion: ecosystem/pickup/pickup-vector-dbs.md, the charter seed, oracle candidates, and first claim registry for this area.
- Evidence pack: ecosystem/pickup/_evidence/vector-dbs.md, the repositories checked against the GitHub API, with dates.
- Verdict source: stack/vector-dbs.md, written under stack/METHOD.md.
Starting a build here?
Begin from the pickup companion for Vector databases: it carries the charter seed, the oracle candidates and how to verify them, and the first claims to register, so a new project starts from evidence instead of from scratch.