# Voyage AI MCP > Voyage AI Embeddings API handles complex data vectorization, letting your agent search by meaning, not just keywords. It generates high-fidelity embeddings for text, code, and images, while also running smart reranking jobs to ensure your retrieval results are surgically precise. ## Overview - **Category:** developer-tools - **Price:** Free - **Tags:** embeddings, rag, rerank, vector-search, multimodal-ai ## Description You need to make sure that when a user asks a question, the system doesn't just match words; it understands the intent behind them. This MCP gives your agent the tools to do that using advanced vectorization and search refinement. Instead of relying on simple keyword matches, you feed complex documents into this service, which converts them into high-dimensional vectors—numerical representations that capture context. If your workflow needs to process millions of records or handle mixed content (like a document with graphs), the batch functions make it scalable. The real power comes when you combine this MCP’s search capabilities with other services; for instance, you can chain this with a messaging MCP and have your agent automatically send a summary of the findings right after retrieval. This entire process runs securely on Vinkius, guaranteeing that every data flow is fully visible through its AI Analytics dashboard. ## Tools ### cancel_batch Stops a batch inference job before it finishes running. ### create_batch Starts a large-scale, asynchronous data processing job. ### create_contextualized_embeddings Generates vector embeddings that retain the meaning of their surrounding document context. ### create_embeddings Creates standard numerical vectors for pure text input. ### create_multimodal_embeddings Generates single vectors from mixed content, like images paired with descriptions. ### delete_file Removes a file that was previously uploaded to the system. ### get_batch Checks the current status and progress of an existing batch job. ### get_file_content Downloads the actual binary or text content of a specific file. ### get_file Retrieves general metadata about a stored file. ### list_batches Shows an overview of all previously created and running batch jobs. ### list_files Lists all files currently stored in the system's repository. ### rerank Scores multiple documents against a given query to find the most relevant context. ### upload_file Uploads a file specifically for use in an asynchronous batch job. ## Prompt Examples **Prompt:** ``` Create embeddings for the text 'What are the benefits of vector search?' using the voyage-4 model. ``` **Response:** ``` I've generated the embeddings for your text using `voyage-4`. The output includes a high-dimensional vector ready for your vector database or similarity search. ``` **Prompt:** ``` Rerank these documents ["Doc A content...", "Doc B content..."] for the query 'AI safety' using rerank-2.5. ``` **Response:** ``` I've reranked the documents. Doc B scored significantly higher (0.92) than Doc A (0.45) for the query 'AI safety'. You should prioritize Doc B in your response. ``` **Prompt:** ``` Start a batch embedding job for the file 'file-987' using the embeddings endpoint. ``` **Response:** ``` The batch job has been initiated successfully. You can monitor its progress using the `get_batch` tool with the ID: `batch_v1_abc123`. ``` ## Capabilities ### Vectorize Text Converts large bodies of text or code into mathematical vectors for semantic search. ### Handle Multimodal Content Creates single, unified vectors from mixed input like images and surrounding text. ### Process Data in Batches Manages large-scale data ingestion by submitting and monitoring asynchronous jobs. ### Improve Search Relevance Takes initial search results and scores them, boosting the most relevant documents to the top for your agent. ## Use Cases ### Technical Manual Search An engineer needs to find a specific fix across 10 years of product manuals. They upload all PDFs using `upload_file`, then run `create_contextualized_embeddings`. When the user asks about 'error code X', the agent uses `rerank` on the top results to pinpoint the exact paragraph, skipping irrelevant sections. ### Legal Document Review A paralegal must review thousands of contracts for mentions of a specific clause. Instead of running 100 separate searches, they use `create_batch` to process all documents at once. They then analyze the results to find every instance of the key phrase. ### Product Catalog Search A user wants to search for a product based on an image and a description. The agent uses `create_multimodal_embeddings` on both inputs, allowing it to match visual intent with textual queries simultaneously. ### Codebase Q&A A developer asks a question about legacy code written in an old language. They use the embedding tools to vectorize the codebase documentation and then retrieve contextually relevant snippets, allowing their agent to answer with high accuracy. ## Benefits - Better search results: Use `rerank` to score documents and ensure your agent only sees the highest-relevance context for its answer. This drastically improves accuracy over basic vector lookups. - Handle massive data loads: If you have millions of records, don't process them synchronously. Use `create_batch` to queue jobs, then check status with `get_batch`, keeping your agent responsive while the background work completes. - Context-aware embeddings: Forget simple text vectors. `create_contextualized_embeddings` embeds chunks while preserving their relationship to the full source document, cutting down on retrieval errors. - Mixed media support: Need to search a manual that contains both text and diagrams? `create_multimodal_embeddings` combines those sources into one searchable vector space. - Full visibility: You can track every step of this process—from initial file upload with `upload_file` to the final scoring—through Vinkius AI Analytics, so nothing happens in the dark. ## How It Works The bottom line is that you manage data lifecycle—from raw file upload to final scored result—all through a sequence of structured API calls. 1. First, upload raw data or file metadata using `upload_file` to prepare it for processing. 2. Next, decide on the embedding type—you might call `create_contextualized_embeddings` if you need document context, or use `create_multimodal_embeddings` for mixed media. 3. Finally, when retrieving information, run `rerank` against your query to score and prioritize the top results before passing them back to the agent. ## Frequently Asked Questions **How do I handle massive volumes of documents with Voyage AI (AI Embeddings API)?** You use the batch tools. First, `upload_file` to stage your data, then call `create_batch`. You can monitor progress and check status using `get_batch` until the job is complete. **What's the difference between `create_embeddings` and `create_contextualized_embeddings`?** Simple embeddings treat text in isolation. Contextualized embeddings use surrounding document information to create a more accurate vector, which is critical for complex documents. **When should I use the `rerank` tool?** Always use it before passing data to the final LLM call. It scores your initial search results against the user's query, guaranteeing you pass the most relevant context possible. **Can this MCP handle images and text together?** Yes. Use `create_multimodal_embeddings` to generate a single vector space that represents both visual data (images) and descriptive text, making them searchable as one unit. **When I need to process a large dataset, what is the proper workflow for using `upload_file`?** You must use `upload_file` first. This action puts the data into the system's queue, making it available for subsequent batch operations like creating embeddings. **If my embedding job fails or stalls, how do I check its status using `get_batch`?** `get_batch` retrieves the current state of a specific batch job. You can use this to confirm if it's running, finished successfully, or if an error occurred. **How do I manage my data retention and clean up temporary assets using `delete_file`?** `delete_file` permanently removes a file from the system. This is crucial for maintaining compliance and keeping your workspace organized after job completion. **Before running any batch operation, how do I see what files are already stored by using `list_files`?** `list_files` retrieves a comprehensive list of every file in the system. This lets you check metadata and confirm your starting data sources before processing. **How does reranking improve my RAG system's accuracy?** By using the `rerank` tool, your agent can take a list of potentially relevant documents and re-score them using a powerful cross-encoder model. This ensures that the most semantically relevant pieces of information are ranked first, providing better context for the LLM to answer queries. **What is the benefit of using contextualized embeddings?** The `create_contextualized_embeddings` tool allows you to embed chunks of text while considering the surrounding content of the same document. This prevents loss of meaning that often happens with standard chunking, leading to much higher retrieval precision. **Can I process images and text in the same vector space?** Yes! With `create_multimodal_embeddings`, you can provide interleaved sequences of text and image URLs. Voyage AI will generate a single embedding that represents the combined semantic meaning, perfect for visual or hybrid search.