voyage-context-3
Voyage-context-3 is an embedding model from Voyage AI that produces chunk-level vectors enriched with global document context, designed for retrieval-augmented generation (RAG) pipelines where long documents are split into smaller chunks.
Publisher review
Voyage-context-3 is an embedding model from Voyage AI that produces chunk-level vectors enriched with global document context, designed for retrieval-augmented generation (RAG) pipelines where long documents are split into smaller chunks. It targets developers and data teams building search and retrieval systems on unstructured text, especially those dealing with documents exceeding typical context windows. The model supports up to 120,000 token context windows, enabling it to process entire documents in a single pass while maintaining fine-grained chunk-level representations. It is a drop-in replacement for standard embeddings in existing RAG workflows, requiring no pipeline changes beyond swapping the model identifier.
Voyage-context-3 uses a nested input format where users provide both the full document and the chunk to be embedded, allowing the model to leverage surrounding context for more accurate chunk representations. It employs Matryoshka learning and quantization-aware training, enabling flexible output dimensions (256, 512, 1024, 2048) and multiple quantization options including int8 and binary formats. This reduces vector database storage costs by up to 99.48% compared to standard float32 embeddings while preserving retrieval accuracy. On benchmarks, it outperforms OpenAI text-embedding-3-large by 14.24% on chunk-level retrieval tasks and 7.89% on document-level retrieval, and improves retrieval accuracy for long documents by 7-23% over models without global context.
In the commercial embedding market, voyage-context-3 competes directly with OpenAI's text-embedding-3-large ($0.13 per 1M tokens, 8,191 token limit) and Cohere's embed-v4 ($0.10 per 1M tokens, 512 token limit). Voyage AI's model is priced at $0.18 per 1M tokens, higher than both competitors, but offers a significantly larger context window and superior retrieval performance on long documents. Its Matryoshka and quantization features also provide storage cost advantages that can offset the higher per-token price in large-scale deployments. The model is accessible via API through Voyage AI's own endpoint and through LiteLLM, which supports the same parameters including input_type, dimensions, and encoding_format.
The primary trade-off is cost: at $0.18 per 1M tokens, voyage-context-3 is more expensive than OpenAI's large model and significantly more than Jina AI's jina-embeddings-v3 ($0.02 per 1M tokens) or open-source alternatives like BGE-large-en-v1.5 (free, self-hosted). The nested input format adds complexity to preprocessing pipelines, requiring both the full document and chunk to be sent in each API call, which increases token consumption. While quantization dramatically reduces storage, it introduces a minor accuracy degradation that may be unacceptable for precision-critical applications. Finally, as a proprietary API, it lacks the flexibility of open-source models for on-premise deployment or fine-tuning on domain-specific data.
How it works
-
Contextualized chunk embeddings
Produces chunk-level vectors that incorporate global document context, improving retrieval accuracy for long documents by 7-23%.
-
Matryoshka learning support
Supports output dimensions of 256, 512, 1024, or 2048, allowing users to trade off storage cost for accuracy.
-
Quantization-aware training
Enables int8 and binary quantization, reducing vectorDB storage costs by up to 99.48% while maintaining retrieval performance.
-
120K token context window
Accepts up to 120,000 tokens per input, enabling whole-document processing without chunking.
-
Drop-in RAG replacement
Seamlessly replaces standard embedding models in existing RAG pipelines without code changes.
-
Nested input format
Accepts both the full document and the chunk in a single API call, using document context to enhance chunk embeddings.
-
API and LiteLLM integration
Available via Voyage AI's API and through LiteLLM, supporting parameters like input_type, dimensions, and encoding_format.
Strengths and trade-offs
Strengths
- Outperforms OpenAI text-embedding-3-large by 14.24% on chunk-level retrieval tasks and 7.89% on document-level retrieval tasks.
- Reduces vector database storage costs by up to 99.48% through int8 and binary quantization while preserving retrieval accuracy.
- Supports a 120,000 token context window, far exceeding OpenAI's 8,191 token limit and Cohere's 512 token limit.
- Improves retrieval accuracy for long documents by 7-23% compared to standard embedding models that lack global context.
Trade-offs
- Priced at $0.18 per 1M tokens, which is 38% more expensive than OpenAI text-embedding-3-large ($0.13 per 1M tokens).
- Requires a nested input format that sends both the full document and chunk per API call, increasing token consumption and preprocessing complexity.
- As a proprietary API, it cannot be self-hosted or fine-tuned on custom data, unlike open-source models such as BGE-large-en-v1.5.
- Quantization to binary or int8 introduces minor accuracy degradation, which may be unacceptable for high-precision retrieval applications.
Pricing context
$0.18 per 1M tokens for API usage. No free tier or self-hosted option available.
Getting started with voyage-context-3
-
Sign up for Voyage AI
Create an account on the Voyage AI website to obtain an API key. This key authenticates your requests to the voyage-context-3 embedding model.
-
Install LiteLLM library
Install the LiteLLM Python library using pip. This library provides a unified interface to call voyage-context-3 and other models, simplifying integration into your RAG pipeline.
-
Set API key as environment variable
Set your Voyage AI API key as the environment variable VOYAGE_API_KEY. This allows LiteLLM to authenticate your requests without hardcoding the key in your scripts.
-
Embed a document chunk
Use LiteLLM's embedding function with model='voyage-3' and input_type='document'. Pass both the full document and the chunk in the nested input format to generate a contextualized embedding.
-
Configure output dimensions
Set the dimensions parameter to 256, 512, 1024, or 2048 to control embedding size. Use Matryoshka learning to balance storage cost and accuracy for your vector database.
Frequently Asked Questions
What is voyage-context-3 and how does it work for RAG?
Voyage-context-3 is an embedding model from Voyage AI that generates chunk-level vectors with global document context. It's designed for RAG pipelines, processing up to 120,000 tokens per input and using a nested format to enhance chunk representations with surrounding document context.
How does voyage-context-3 compare to OpenAI text-embedding-3-large?
Voyage-context-3 outperforms OpenAI text-embedding-3-large by 14.24% on chunk-level retrieval and 7.89% on document-level tasks. It offers a 120,000 token context window versus OpenAI's 8,191 limit, but costs $0.18 per 1M tokens compared to OpenAI's $0.13.
What is the nested input format in voyage-context-3?
The nested input format requires sending both the full document and the specific chunk in each API call. This allows the model to use global document context to produce more accurate chunk embeddings, though it increases token consumption and preprocessing complexity compared to standard embedding models.
How does voyage-context-3 reduce vector database storage costs?
Voyage-context-3 uses Matryoshka learning and quantization-aware training, supporting output dimensions from 256 to 2048 and int8 or binary formats. This can reduce vector database storage costs by up to 99.48% compared to standard float32 embeddings while maintaining retrieval accuracy.
What are the main weaknesses of voyage-context-3?
Voyage-context-3 costs $0.18 per 1M tokens, 38% more than OpenAI's large model. Its nested input format increases token usage and pipeline complexity. As a proprietary API, it cannot be self-hosted or fine-tuned, and quantization may cause minor accuracy loss for precision-critical tasks.
How can I integrate voyage-context-3 into my existing RAG pipeline?
Voyage-context-3 is a drop-in replacement for standard embedding models, requiring no code changes beyond swapping the model identifier. It's accessible via Voyage AI's API or through LiteLLM, supporting parameters like input_type, dimensions, and encoding_format for seamless integration.
Alternatives
How voyage-context-3 compares
Direct head-to-head against 2 competitors. Picked by 7wData.
voyage-context-3
- Pricing
- $0.18 per 1M tokens for API usage. No free tier or self-hosted option available.
- Target
- Voyage-context-3 is an embedding model from Voyage AI that produces chunk-level vectors enriched with global document context, designed for retrieval-augmented generation (RAG) pipelines where long
- Strength
- Outperforms OpenAI text-embedding-3-large by 14.24% on chunk-level retrieval tasks and 7.89% on document-level retrieval tasks.
- Watch for
- Priced at $0.18 per 1M tokens, which is 38% more expensive than OpenAI text-embedding-3-large ($0.13 per 1M tokens).
OpenAI text-embedding-3-large
- Pricing
- $0.13/1M tokens
- Target
- General-purpose embeddings
- Deployment
- API-only
- Strength
- 3072-dimension default embeddings
- Watch for
- No Matryoshka/quantization options
Cohere Embed v3
- Pricing
- $0.50/1M tokens
- Target
- English-language retrieval
- Deployment
- API-only
- Strength
- 1024-dimension embeddings
- Watch for
- 512-token context window limit
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.