# Vertex AI Atlas Vector Search > Source-anchored research for Vertex AI Atlas Vector Search. Run status: COMPLETED. Parent: [MongoDB Atlas on GCP](https://llms-explorer.com/tree/mongodb-atlas-on-gcp/) · 3 facets · 16 facts · page: https://llms-explorer.com/tree/vertex-ai-atlas-vector-search/ ## How it works - The multi-agent reference can fall back to case-insensitive regular-expression search when the index, stored embeddings, or aggregation is unavailable. Its method field distinguishes vector_search from regex_fallback; a successful response alone does not demonstrate semantic retrieval. (low confidence) — [source](https://github.com/mongodb/multi-agent-using-mongodb-for-hyperscaler) - Google’s lifecycle page lists April 1, 2027 as the retirement date for text-embedding-005. This is a scheduled service lifecycle date and should be rechecked when selecting the model for an Atlas integration. (low confidence) (verified as of 2026-10-23) — [source](https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions) ## Comparisons and alternatives - SmartSoft’s drilling-incident prototype reports combining Vertex AI/Gemini, Atlas vector indexing, and Cloud Run. Its authors identify embedding-size management, latency versus answer accuracy, and cloud integration as implementation challenges; the write-up supplies no controlled performance benchmark. (low confidence) — [source](https://devpost.com/software/smart-digital-drilling-incident-investigator) ## Facts and statements - MongoDB’s documented Vertex AI integration stores document embeddings in Atlas and uses Atlas vector retrieval with a Vertex AI chat model. Its linked Qwiklab repository implements the demonstration with Streamlit and LangChain. (high confidence) — [source](https://www.mongodb.com/docs/atlas/ai-integrations/google-vertex-ai.md) - The documented PDF tutorial creates vector_index on vertexaiApp.chat-vec with 768 dimensions and requires an Atlas IP access-list entry plus a Google Cloud project with the Vertex AI API enabled. These are tutorial settings, not universal requirements for every embedding model. (low confidence) — [source](https://www.mongodb.com/docs/atlas/ai-integrations/google-vertex-ai.md) - MongoDB’s Vertex AI Extensions tutorial implements natural-language database operations through an OpenAPI specification and a Cloud Run function; it is a distinct integration path from the vector-retrieval example. (medium confidence) — [source](https://www.mongodb.com/docs/atlas/ai-integrations/google-vertex-ai/extensions.md) - An independent Spring Boot example combines Vertex AI embeddings with Atlas movie-vector search, demonstrating that the integration is not confined to the Python Streamlit sample. (medium confidence) — [source](https://github.com/Eainde/spring-vertexai-vector-search) - The MongoDB multi-agent reference describes embedding stored documents and incoming queries with the same Vertex model, followed by an Atlas $vectorSearch aggregation. (low confidence) — [source](https://github.com/mongodb/multi-agent-using-mongodb-for-hyperscaler) - LangChain issue #33872 records an ImportError for _DirectlyInjectedToolArg in an environment including langchain-google-vertexai and langchain-mongodb. This is historical dependency-compatibility evidence, not proof of a current Vertex-to-Atlas service defect. (low confidence) — [source](https://github.com/langchain-ai/langchain/issues/33872) - The multi-agent sample names gemini-2.5-flash as its default model, but Google’s lifecycle page lists its retirement date as October 20, 2026. Reusing that sample requires checking and replacing retired model identifiers. (medium confidence) (verified as of 2026-10-23) — [source](https://github.com/mongodb/multi-agent-using-mongodb-for-hyperscaler) - Google’s lifecycle table lists April 1, 2027 as the retirement date for text-embedding-004 and text-embedding-005, making embedding-model lifecycle planning relevant to Vertex-backed Atlas implementations. (low confidence) (verified as of 2026-10-23) — [source](https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions) - Spring AI provides a composable integration path: VertexAiTextEmbeddingModel implements EmbeddingModel, and MongoDBAtlasVectorStore accepts an EmbeddingModel. Combining these documented interfaces allows Vertex AI embeddings to feed an Atlas vector store; this is an inference from the APIs, not a separately benchmarked deployment. (medium confidence) — [source](https://docs.spring.io/spring-ai/reference/api/embeddings/vertexai-embeddings-text.html) - The MongoDB tutorial specifies a 768-dimensional vector_index on vertexaiApp.chat-vec. Do not treat 768 as universal: Google documents a default 3072-dimensional output for gemini-embedding-001 and an output_dimensionality parameter. An application changing models must align its actual embedding output and Atlas index configuration. (medium confidence) (verified as of 2026-10-22) — [source](https://www.mongodb.com/docs/atlas/ai-integrations/google-vertex-ai.md) - Google’s lifecycle table records textembedding-gecko@001 and @002 as retired on April 21, 2025, and @003 as retired on May 24, 2025. Historical Gecko-based examples therefore require model-lifecycle review before reuse. (low confidence) (verified as of 2026-10-23) — [source](https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions) - Atlas is available on Azure, AWS, and Google Cloud, while the reviewed MongoDB Vertex AI tutorial places application compute on a Google Cloud VM. Applying this example to Azure-hosted Atlas is a cross-cloud design inference; these sources do not establish Azure-specific connectivity, latency, or egress costs. (medium confidence) — [source](https://www.mongodb.com/docs/atlas/ai-integrations/google-vertex-ai.md) - The Spring AI MongoDB Atlas reference requires a configured vector-search index and documents schema initialization as opt-in through spring.ai.vectorstore.mongodb.initialize-schema=true. Applications using Vertex AI embeddings through Spring AI must provision the index or enable initialization. (low confidence) (verified as of 2026-10-23) — [source](https://docs.spring.io/spring-ai/reference/api/vectordbs/mongodb.html)