Skip to content

SYSTEM Cited by 3 sources

Faiss

Faiss — Facebook AI Similarity Search — is Meta's open-source library for efficient similarity search and clustering of dense vectors. First disclosed in the 2017 Meta Engineering post; source at github.com/facebookresearch/faiss. One of the most widely deployed vector-similarity-search libraries in the industry; used inside Meta and extensively in third-party production stacks.

Yelp face — in-memory vectorstore for customer-support RAG (2026-05-27)

Yelp's sources/2026-05-27-yelp-beyond-the-menu-tree-how-yelp-built-a-smarter-customer-success-chatbot|2026-05-27 LLM-Assisted CS Chatbot post canonicalises a distinct production altitude for Faiss on the wiki: in-memory similarity search inside a customer-support chatbot container. Distinct from Meta Groups Search's L2 ANN substrate altitude (production scale-out across an item index of group posts) and SilverTorch's Faiss-GPU baseline framing (recsys retrieval over 80M items); Yelp's altitude is small-corpus, single-process, in-container.

Substrate disclosure (verbatim):

"We use Text-embedding-ada-002 to construct our vectorstore. Each individual text segment derived from the metadata is converted into a 1536-dimension unit vector. The entire vectorstore is highly compact, measuring around 8 megabytes. […] We utilize FAISS search for smart indexing and quantization to quickly find the closest vectors to the query."

Operational shape:

Property Value
Corpus ~370 Yelp Support Center articles, ~5 metadata segments per article ≈ ~1,850 vectors
Embedding OpenAI text-embedding-ada-002, 1,536 dim
Quantized size ~8 MB (FAISS "smart indexing and quantization")
Residency In-process container memory — no remote vector DB
Build cadence Every container start during health check
Disclosed quality ~94% recall@5 on Yelp's evaluation dataset

The post does not disclose the specific FAISS index type (IVF / HNSW / IVFPQ / OPQ) or quantisation scheme; "smart indexing and quantization" is the only altitude given. The 8 MB number for ~1,850 × 1,536-dim vectors implies aggressive quantisation (raw ≈ ~11 MB unindexed).

This is the wiki's first canonical Faiss instance at customer-support / chatbot RAG altitude — sibling to:

  • Meta Groups Search L2 (2026-04-21) — Faiss as production ANN over a precomputed index of group-post embeddings; query-side dense-retrieval substrate for the Groups Scoped Search L2 ranker.
  • SilverTorch baseline (2026-05-26) — Faiss-GPU as the per-service ANN-only baseline that SilverTorch's fused Int8 ANN primitive replaces inside Meta's recsys retrieval surfaces.
  • Yelp CS Chatbot (2026-05-27) — Faiss as single-container in-memory ANN library for a small (~8 MB) RAG vectorstore, reloaded daily via S3 batch CSV pipeline, rebuilt at container-start (patterns/in-memory-vectorstore-loaded-at-container-start).

The three altitudes are non-overlapping production shapes — Faiss's API surface accommodates billion-scale Meta-internal recsys work, mid-scale Meta-search-product workloads, and ≤10-MB-corpus single-container chatbot deployments without a different library choice.

SilverTorch face — Faiss-GPU as the baseline (2026-05-26)

In Meta's 2026-05-26 SilverTorch post (Source: sources/2026-05-26-meta-silvertorch-index-as-model-a-new-retrieval-paradigm-for-recommendation-systems), Faiss-GPU is named as the strong same-substrate baseline SilverTorch's fused Int8 ANN primitive replaces inside Meta's recsys retrieval surfaces. Disclosed comparison on an 80M-item production retrieval workload:

Metric FAISS-CPU FAISS-GPU SilverTorch
Compute-cost efficiency vs CPU baseline baseline 5.9× 20.9× (13.35× with reranking)
Maximum top-k unlimited (slow) 2,048 hundreds of thousands
Neural reranking ✗ ✗ ✓
Multi-task scoring ✗ ✗ ✓

Per-primitive: SilverTorch's fused Int8 ANN kernel is 2.2–14.7× faster than Faiss-GPU.

The structural argument: Faiss is "built to find nearby items" — a clean per-service ANN-only API — but recommendation systems "often need to pull back a much larger pool of candidates so later stages can make better relevance decisions." The ceiling on Faiss-GPU isn't a library defect; it's a fit-with-Index-as-Model question. Faiss-GPU continues to be the right choice when ANN search is the workload boundary and the next stages live in separate services. SilverTorch is the right choice when ANN should compose with eligibility filtering + neural reranking + multi-task scoring inside one model graph. See the patterns/gpu-native-retrieval-primitive-redesign pattern for the design philosophy.

Important scope note: SilverTorch supersedes Faiss-GPU inside the recsys retrieval surfaces it now powers (Facebook + Instagram + Threads feed / video) — not Faiss-the-library across all Meta search/retrieval workloads. Faiss continues as the production ANN substrate for systems/meta-groups-scoped-search (per the 2026-04-21 post) and is a Meta-stewarded open-source library used widely outside Meta.

In the 2026-04-21 Meta post, Faiss is named as the production approximate-nearest-neighbor substrate for the semantic-retrieval arm of Facebook Groups Scoped Search:

"We then perform an approximate nearest neighbor (ANN) search over a precomputed Faiss vector index of group posts. This enables the retrieval of content based on high-dimensional conceptual similarity, regardless of keyword overlap."

Query embeddings from SSR are looked up against a precomputed Faiss index of group-post embeddings; nearest neighbors feed the L2 MTML ranker with cosine-similarity scores.

Capabilities (industry-documented)

Not itself a primary focus of the 2026-04-21 post, but widely documented:

  • Multiple ANN index types — IVF (inverted-file), HNSW (graph-based), PQ (product quantisation), OPQ, and composites.
  • CPU and GPU implementations.
  • Exact search and various approximation levers.
  • Billion-scale datasets on a single machine; larger-scale via sharding.

The 2026-04-21 post does not disclose which Faiss index type (IVF / HNSW / PQ / OPQ) Meta uses for Groups scoped search, nor the vector dimensionality, shard layout, or recall/latency targets.

Sibling systems on the wiki

Other vector-index substrates surface on the wiki in adjacent contexts: DiskANN, SPANN, SPFresh, HNSW. Faiss is Meta's default production choice for this class of workload.

Seen in

Last updated · 542 distilled / 1,571 read