PINECONE-SPEC

Pinecone managed vector database SaaS leader.

Définition

Pinecone features : (1) Vector storage : embedding vectors (768 dim ada-002 OpenAI, 1536 dim text-embedding-3-large OpenAI, 384 dim sentence-transformers MiniLM, 4096 dim Mistral, etc.). (2) Index types : index HNSW (Hierarchical Navigable Small World) + IVF-PQ (Inverted File Product Quantization), Pinecone proprietary optimizations. (3) Metadata filtering : key-value metadata associated each vector, filter at query time (category, source, date, etc.). (4) Namespaces : multi-tenant per index isolation. (5) Hybrid search : combine sparse vectors (BM25, SPLADE) + dense vectors, single fused score. (6) Pinecone Inference : embedded models hosting (host embedding model alongside vector DB, single API call generate embedding + query). Pricing Serverless : $0.085/M read units + $4.00/M write units + $0.33/GB-month storage (us-east-1), ~10x cheaper than Pod-Based traditional. SDKs : Python, Node.js, Java, Go, .NET. Integration : LangChain, LlamaIndex, Vercel AI SDK, OpenAI Assistants API. Customers : Notion AI, Shopify, Brex, Cohere, You.com, Glean.

Origine

Pinecone fondee 2019 par Edo Liberty (ex-Yahoo Research, AWS) + Greg Levitan ; Series A $10M 2021 ; Series B $100M 2023 (Andreessen Horowitz lead, $750M valuation) ; Pinecone Serverless launched mars 2024.

Exemple en contexte

Notion AI utilise Pinecone Serverless pour RAG search ses ~100M+ user pages workspace : OpenAI text-embedding-3-large embeddings stockees + searchable via Pinecone, user query 'find my product roadmap docs' returns top-10 semantically similar pages in <100ms.

Termes liés

• Weaviate — concurrent OSS.