MARQO-AI

Marqo end-to-end multimodal vector search AI.

Définition

Marqo features : (1) End-to-end : embedding + indexing + search single Marqo deployment, no separate embedding pipeline (unlike Pinecone, Qdrant which expect pre-computed embeddings). (2) Built-in models : CLIP variants (ViT-B/32, ViT-L/14), OpenCLIP, Sentence Transformers, FastText, etc., automatic embedding generation at index time + query time. (3) Multimodal : add documents with text + image + video URLs, embeddings computed automatically per modality, semantic search across modalities (search 'red shoes' returns both text descriptions + product images). (4) Index : HNSW (Marqo Lucene underlying) + Faiss + Open-source vector indexes. (5) Filtering : pre-filtering scalar fields integrated HNSW traversal. (6) Tensor search : sub-document attention multi-vector per document (entire document split chunks, each chunk indexed separately, search returns best matching chunks). (7) Distributed : Marqo Cluster mode horizontal scaling. SDKs : Python primary, JavaScript emerging. Customers : ~5000+ Marqo Cloud customers 2024, AI startups + retail e-commerce semantic search.

Origine

Marqo fondee 2021 a Sydney par Tom Hamer + Jesse Clark ; Seed $5M 2022 ; Series A 2023 ; ~5000+ customers Marqo Cloud 2024.

Exemple en contexte

Fashion e-commerce startup uses Marqo for visual + text product search : 'add documents' API with product image URLs + descriptions text, Marqo OpenCLIP model auto-embeds, customer search 'casual summer dress' returns matched products combining text + visual similarity in single API call ~50ms latency.

Termes liés

• Chroma — alternative developer-first.