Embedding models and rerankers for semantic search and RAG pipelines
Voyage AI documents state-of-the-art embedding models and rerankers at docs.voyageai.com for retrieval-augmented generation and semantic search. The embeddings API at POST https://api.voyageai.com/v1/embeddings converts text into dense vectors; the Python `voyageai` package exposes `voyageai.Client.embed()` with models such as voyage-4-large, voyage-4, voyage-4-lite, voyage-code-3, and domain-specific variants per docs.voyageai.com/docs/embeddings. Docs describe `input_type` query vs document prompts for retrieval tasks, optional `output_dimension` (256–2048), and `output_dtype` quantization (float, int8, binary). Rerankers score query–document relevance to refine initial retrieval results. Voyage AI is part of MongoDB's AI stack (MongoDB docs reference Voyage embeddings) and integrates with vector stores and LLMs as modular RAG components.
Use cases
- First-stage vector retrieval before LLM generation in RAG chatbots
- Reranking BM25 or embedding hits for higher-precision result sets
- Domain-specific retrieval with voyage-code-3 or voyage-law-2 models
- Pair with Pinecone, Weaviate, or Vespa entries when comparing embedders
- Cost/latency tuning via voyage-4-lite and output_dtype quantization
Key features
- Text embedding API with query/document input_type tuning for retrieval
- Reranker models scoring query–document relevance pairs
- Configurable output dimensions and quantization dtypes
- Python voyageai client and REST POST /v1/embeddings endpoint
- Multilingual semantic similarity per Voyage docs
Who Is It For?
- Developers building RAG pipelines needing high-accuracy embeddings
- Search teams adding rerankers after initial lexical or vector retrieval
- Teams evaluating embedders alongside Cohere or OpenAI embedding APIs
Frequently Asked Questions
- Does Voyage AI host vector storage?
- Voyage docs describe embedding and reranking APIs; vector storage is typically handled by separate vector databases integrated into your RAG stack.
- What is input_type query vs document?
- Docs say Voyage prepends retrieval-tuned prompts when input_type is query or document for search/RAG use cases.
- How does Voyage relate to MongoDB?
- MongoDB documentation references Voyage AI embeddings as part of MongoDB's AI ecosystem; verify current integration paths in MongoDB and Voyage docs.
Related
Related
3 Indexed items
Chroma
Chroma documents an open-source embedding database at docs.trychroma.com for storing and querying vectors, metadata, and full-text fields in Python and JavaScript clients. Official guides cover ephemeral in-memory collections, persistent local storage, self-hosted server deployments, and Chroma Cloud at trychroma.com with authentication tokens. The docs describe collection CRUD, `add`/`query`/`get`/`update`/`delete` APIs, embedding functions (default and third-party), hybrid search, and multitenancy patterns for RAG and agent memory workloads per the documentation index.
Jina AI
Jina AI documents Search Foundation APIs at docs.jina.ai for RAG and agent pipelines. The Reader API at POST https://r.jina.ai/ converts a single URL into LLM-friendly markdown or JSON with optional headers such as X-Engine (browser/direct), X-Return-Format, and X-Respond-With readerlm-v2 per docs. The Search API at POST https://s.jina.ai/ returns SERP-style web results optimized for downstream LLM use with query parameters q, gl, hl, num, and page. Embeddings API at POST https://api.jina.ai/v1/embeddings supports models including jina-embeddings-v5-text-small, jina-embeddings-v5-text-nano, jina-embeddings-v4, jina-embeddings-v3, and jina-clip-v2 with task types such as retrieval.query and retrieval.passage. Reranker API at POST https://api.jina.ai/v1/rerank includes jina-reranker-v3 for refining retrieval hits. EU endpoints eu.r.jina.ai and eu.s.jina.ai keep processing in EU jurisdiction per docs.
Milvus
Milvus documents a high-performance vector database at milvus.io/docs for storing, indexing, and searching embedding vectors with metadata filtering and hybrid search. Deployment options include Milvus Lite (`pip install pymilvus` for notebooks/edge), Milvus Standalone (single Docker image), and Milvus Distributed on Kubernetes per milvus.io/docs/v2.6.x/install-overview. Official SDKs include PyMilvus, Go, Java, Node.js, and C#; Zilliz Cloud offers managed Milvus. Architecture separates access, coordinator, worker, and storage layers with object storage backends (MinIO, S3, Azure Blob) per milvus.io/docs/architecture_overview.