Local embeddings from Ollama into MongoDB Atlas Vector Search
Parent: Ollama deployment, cloud, security and operations · Published reference · snapshot 2026-10-14 · skill ollama-expert/references/ollama-ops-platform.md
Also known as: Local Ollama embeddings into MongoDB Atlas Vector Search
↓ Facts as markdownall context files
Everything the scanned sources say about Local Ollama embeddings into MongoDB Atlas Vector Search: 22 source-anchored units across 4 sources, grouped by facet.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Definitions
- modelName is the name you use in the autoEmbed index. [source]
Structure and components
- The catalog embedding-service-configs.yml is installed next to the mongot binary and already contains two ready-to-use local models pointed at Ollama: bge-m3 at 1024 dimensions and nomic-embed-text at 768 dimensions. [source]
How-to and procedures
Measurements and reference values
- Index quantization: scalar cuts vector size 4x and mongot RAM about 3.75x; binary cuts vector size 32x and RAM about 24x (HNSW graph not compressed). MongoDB recommends it for collections over ~100,000 vectors, binary with rescoring for most models, scalar for low-dimension non-QAT models. Query vectors are quantized automatically. [source]
Problems, failure modes and limitations
- nomic-embed-text is asymmetric: Percona's config sets documentPrefix 'search_document: ' and queryPrefix 'search_query: '. Ollama passes input through unchanged, so omitting prefixes gives no error but degraded retrieval. [source]
Facts and statements
- mongot autoEmbed can be pointed at Ollama via embeddingProvider OPENAI_COMPATIBLE, providerEndpoint http://localhost:11434/v1/embeddings, modelName nomic-embed-text, outputDimensions 768, empty credentials; catalog file embedding-service-configs.yml (custom via embedding.modelConfigFile, restart mongot). This is Percona's documentation. [source]
- MongoDB's official local-RAG tutorial uses Ollama: it pulls nomic-embed-text for embeddings and Mistral 7B as the generator, with the Atlas CLI (v1.14.3+) and a local Atlas deployment. This applies to the .NET, Go and Java variants of the tutorial; the Node.js variant uses mxbai-embed-large-v1 and GPT4All, and the Python variant uses atlas-local-lib-py without the CLI. [source]
- Ollama's /api/embed returns L2-normalized (unit-length) vectors and accepts a list in input for batching, so dotProduct (which requires unit-length vectors) is valid with Ollama output; cosine also works. [source]
- nomic-embed-text produces 768-dimension vectors, so the index needs type vector, numDimensions 768, path to the embedding field, similarity cosine; a wrong index name returns no results silently. [source]
- numDimensions is required and max 8192; binary quantization needs numDimensions to be a multiple of 8; cosine rejects zero-magnitude vectors. Query vectors must have the same dimension and come from the same model as stored vectors, so a model change means re-embedding and reindexing. [source]
- I create a vector search index (Ollama nomic-embed-text, like BERT, has 768 dimensions): [source]
- Pre-quantized ingestion: embeddings can be stored as BSON binData (float32, int8, int1); int1 supports euclidean only and needs numDimensions multiple of 8; PyMongo 4.10+, Node 6.11+ etc. MongoDB lists nomic-embed-text-v1.5 and mxbai-embed-large-v1 among models that output byte vectors. [source]
- Build a Local RAG Implementation with MongoDB Vector Search [source]
- db.eurovision.createSearchIndex( "vectorSearchOnLyrics", "vectorSearch", { fields: [ { type: "vector", numDimensions: 768, path: "embedding", similarity: "cosine" } ] } ) [source]
- Local Atlas deployments are intended for testing only; MongoDB says to deploy a cluster for production. A MongoDB Community or Enterprise cluster with Search and Vector Search installed is the alternative. [source]
- The following example uses nomic-embed-text, which produces 768-dimensional vectors in the Percona model catalog. [source]
- MongoDB Vector Search Index, With Local Atlas and Ollama [source]
- This article discusses implementing vector search in a local MongoDB Atlas setup using Ollama and the Eurovision song lyrics dataset. We populated the database with vectorized data embedded in the documents and created a vector search index for retrieving songs based on semantic similarity. [source]
- Follow the Install and configure mongot procedure to set up Percona Search for MongoDB. Configure PSMDB as a replica set, point the search parameters to mongot, and create a user with the searchCoordinator role. [source]
- You can use Ollama to generate embeddings locally for Percona Search for MongoDB. [source]
- Local Atlas deployments are intended for testing only. For production environments, deploy a cluster. [source]
Related concepts
- ollama — is a related of Local Ollama embeddings into MongoDB Atlas Vector Search
- nomic-embed-text — is a instance of Local Ollama embeddings into MongoDB Atlas Vector Search
- numDimensions — is a part of Local Ollama embeddings into MongoDB Atlas Vector Search
- percona — is a instance of Local Ollama embeddings into MongoDB Atlas Vector Search
- mongot — is a part of Local Ollama embeddings into MongoDB Atlas Vector Search
- l2-normalized — is a measure of Local Ollama embeddings into MongoDB Atlas Vector Search
- search_document: — is a part of Local Ollama embeddings into MongoDB Atlas Vector Search
- vector quantization — is a hyponym of Local Ollama embeddings into MongoDB Atlas Vector Search
- 11434 — is a related of Local Ollama embeddings into MongoDB Atlas Vector Search
Children
- No children recorded.