Run Nomic Embed Privately | 8K Context RAG Embeddings

Why Nomic Embed Is the Long-Context Default

Most embedding models in the Ollama catalog were designed for sentence or paragraph-level inputs. Their context windows top out at 256 to 512 tokens, which means any document longer than a short paragraph must be chunked before embedding — and the chunking strategy becomes a critical, difficult-to-tune variable in retrieval quality.

Nomic Embed Text was built to a different specification. Its 8,192-token context window means a full page of dense text, a multi-section technical article, or a complete short contract can be embedded as a single unit. The retrieval pipeline becomes simpler, context is preserved across boundaries that chunking would otherwise split, and the semantic representation of a document reflects its full meaning rather than fragment-level patterns.

Nomic Embed v2 MOE extends this further with a Mixture-of-Experts architecture that improves quality across both English and multilingual retrieval tasks while remaining efficient enough to run on CPU hardware. For organizations that have outgrown All-MiniLM's token limit or need more semantic depth than a small embedding model provides, Nomic Embed is consistently one of the first upgrades the RAG community reaches for.

8,192

Max Token Context

768

Default Dimensions

64→768

Matryoshka Dim Range

Apache 2.0

License

Nomic Embed Model Family on Ollama

Nomic Embed Text v1.5

Recommended — Standard

137M parameters, 768 dimensions, 8192-token context. Matryoshka training allows truncation to 64, 128, 256, or 512 dimensions with graceful quality degradation. Strong MTEB score of 62.4. The established workhorse for long-document private RAG.

ollama pull nomic-embed-text RAM: ~270MB · Dims: 768 (flexible)

Nomic Embed v2 MOE

Recommended — Best Quality

475M parameters, Mixture-of-Experts architecture, 8192-token context, improved multilingual coverage. Stronger MTEB performance than v1.5 with better handling of non-English text. The upgrade path when v1.5 retrieval quality is the limiting factor.

ollama pull nomic-embed-text:v2 RAM: ~950MB · Dims: 768

Start with Nomic Embed Text v1.5. It handles the overwhelming majority of long-document RAG use cases with a minimal footprint — under 300MB RAM means it coexists with any generation model without competition. Upgrade to v2 MOE when retrieval quality testing on your specific corpus reveals gaps, or when multilingual document coverage is a requirement.

Who This Model Family is For

👤 Personal use

Nomic Embed is useful if you want to build a private semantic search index over your own notes, PDFs, research documents, or personal knowledge base. It does not chat or generate answers on its own — it converts text into vectors so another model (such as Llama or Qwen) can retrieve the most relevant passages. If you're building a personal "second brain" or local RAG system, Nomic Embed is one of the strongest fully open options available.

🏢 Business / team use

Teams use Nomic Embed as the retrieval backbone for internal knowledge bases, contract search, support archives, and document Q&A systems. Its 8192-token context window allows you to embed longer documents with less aggressive chunking, preserving more meaning. Because the weights are fully open under Apache 2.0, sensitive content never needs to leave your infrastructure — unlike sending queries to closed embedding APIs.

⚙️ Developer & technical use

Developers building RAG pipelines will appreciate Nomic Embed v1.5 for its Matryoshka Representation Learning: you can truncate the 768-dimensional embeddings down to 512, 256, 128, or 64 dimensions with graceful quality loss. This gives you fine-grained control over vector database storage costs. It also supports task-specific prefixes (search_query:, search_document:, etc.), allowing you to optimize embeddings for the specific retrieval task.

🔬 Data science & research

Researchers working on retrieval, clustering, or topic modeling benefit from Nomic Embed being one of the most reproducible open embedding models — training data, code, and weights are all publicly available. It delivers competitive MTEB retrieval scores while offering a long context window and flexible embedding dimensions, making it a strong baseline for ablation studies and cost-quality tradeoff research.

Licensing Information

Nomic Embed v1 and v1.5 are released under the Apache 2.0 License, which permits commercial use, modification, redistribution, and private deployment without royalty obligations or usage caps. This makes it one of the cleanest choices for commercial RAG and search products.

Official model card:

nomic-ai/nomic-embed-text-v1.5 on Hugging Face

Disclaimer: NoCloudGPT provides this licensing summary as a convenience reference only. It does not constitute legal advice. License terms can change. Always review the official model card and consult your legal counsel before deploying any model in a commercial or regulated context.

Available Model Sizes

Model Size	Ollama Tag	Recommended NoCloudGPT Tier	Notes
Nomic Embed v1.5 (137M params)	nomic-embed-text	Starter or higher	Default recommendation. 8192-token context, Matryoshka dimensions (768→64), strong MTEB performance.	Get quote →
Nomic Embed v1 (137M params)	nomic-embed-text:v1	Starter or higher	Original release without Matryoshka. Use only if matching an existing v1 index.	Get quote →

Note: Embedding models are lightweight compared to chat models. The performance bottleneck in a RAG system is almost always the generation model (Llama, Qwen, Mistral, etc.), not the embedder. Nomic Embed runs efficiently even on modest hardware.

What Makes Nomic Embed Different

📄

8K Token Context Window

Embed full pages, complete policy documents, and multi-section reports as single units. Eliminates the context fragmentation problem that plagues short-context embeddings on document-length inputs.

🎯

Matryoshka Dimensions

Nomic Embed is trained with Matryoshka Representation Learning, meaning the first N dimensions of any embedding are themselves a valid, high-quality embedding. Truncate to 256 dimensions for 3x storage savings with less than 5% quality loss.

🔓

Fully Open Training

Nomic AI published the complete training data, methodology, and codebase. Unlike many open-weight models with closed training pipelines, Nomic Embed's provenance is fully auditable — critical for organizations with AI governance requirements.

🧩

MOE Architecture (v2)

Nomic Embed v2 uses a Mixture-of-Experts design where different expert networks handle different input types. This gives v2 stronger performance on diverse document corpora where a single dense model underperforms on vocabulary outliers.

⚡

CPU-Efficient Inference

Despite the 8K context capability, Nomic Embed v1.5 runs efficiently on CPU hardware. The 137M parameter count keeps RAM footprint under 300MB and inference latency acceptable for real-time RAG queries.

🔗

Universal Framework Support

Nomic Embed is a first-class citizen in LangChain, LlamaIndex, Haystack, and OpenWebUI's RAG pipeline. Pull it via Ollama and configure it as the embedding provider in minutes with no framework-specific adaptation.

Understanding Matryoshka Embeddings in Practice

Matryoshka Representation Learning (MRL) is a training technique that optimizes embedding quality at multiple dimension sizes simultaneously. For Nomic Embed, this means the 768-dimensional vector it produces can be truncated to smaller sizes without retraining — and each truncated version is still a well-trained embedding at that dimension count.

The practical implications for RAG deployment are significant. A standard embedding pipeline stores 768-dimensional vectors. With Nomic Embed's Matryoshka property, you can choose to store 256-dimensional vectors instead — one-third the storage cost — and accept a retrieval quality trade-off that, for most use cases, is genuinely small. On typical enterprise document corpora, the quality difference between 256 and 768 dimensions with Nomic Embed is less than 3% on recall metrics.

For organizations with very large document corpora — millions of documents, petabyte-scale vector stores — the storage and index query cost reduction from Matryoshka truncation is a material operational saving. NoCloudGPT can help you measure the quality-versus-storage trade-off on your specific corpus before committing to a dimension size.

Matryoshka Dimension Selection Guide

# Nomic Embed Text v1.5 — Matryoshka dimension options: # 768 dims → Full quality, standard storage (~3KB per vector) # 512 dims → ~2% quality loss, 33% storage reduction # 256 dims → ~3-4% quality loss, 67% storage reduction # 128 dims → ~6-8% quality loss, 83% storage reduction # 64 dims → Significant quality loss, prototype/cache use only # Recommended for most deployments: 512 dims (best trade-off)

Truncating Embeddings in Python

import numpy as np def truncate_embedding(embedding, target_dim): # Nomic Embed MRL — truncate and re-normalize truncated = np.array(embedding[:target_dim]) normalized = truncated / np.linalg.norm(truncated) return normalized.tolist() # Use with cosine similarity search # Example: truncate 768-dim to 256-dim compact_vec = truncate_embedding(full_embedding, 256)

Long Document Chunking Strategy

# With 8K context — use larger, semantically complete chunks # Recommended for Nomic Embed: chunk_size = 1500 # tokens — embed whole sections/pages chunk_overlap = 150 # 10% overlap at section boundaries # Prefer splitting at: paragraph breaks, section headers, list ends # Avoid splitting: mid-sentence, mid-table, mid-code-block # Result: fewer chunks, simpler index, better context preservation

OpenWebUI RAG Configuration

# OpenWebUI Settings → Documents → Embedding Model # Set to: nomic-embed-text (via Ollama) # Chunk size: 1500 tokens # Chunk overlap: 150 tokens # Top K retrieval: 5 (higher quality than MiniLM needs) # Similarity threshold: 0.70 # This config handles full-page PDFs without fragmentation

Hardware Requirements

Model	Min RAM	Recommended Instance	CPU Throughput	Notes
Nomic Embed v1.5	2 GB	Lightsail 2–4GB	60–120 docs/s	Runs alongside generation model on 4GB+
Nomic Embed v2 MOE	4 GB	Lightsail 4–8GB	20–50 docs/s	Higher quality, more RAM — use dedicated instance at scale
v1.5 + LLM (same server)	6 GB	Lightsail 8GB	Unaffected by LLM	Full RAG stack — embed + generate on one instance
v2 MOE + LLM (same server)	10 GB	GPU instance / on-premise	Varies	Recommended to separate embedding and generation

Deployment with Ollama

        # Pull Nomic Embed Text v1.5

        ollama pull nomic-embed-text

        # Pull Nomic Embed v2 MOE

        ollama pull nomic-embed-text:v2

        # Generate embedding via API

        curl http://localhost:11434/api/embed \

          -d '{

            "model": "nomic-embed-text",

            "input": "search_document: Full page of enterprise documentation..."

          }'

        # Note: Nomic Embed uses task prefixes for best results

        # "search_document:" prefix for documents being indexed

        # "search_query:" prefix for user queries at retrieval time

        # LangChain integration

        from langchain_community.embeddings import OllamaEmbeddings

        embeddings = OllamaEmbeddings(

          model="nomic-embed-text",

          show_progress=True

        )

        # LlamaIndex with Nomic Embed

        from llama_index.embeddings.ollama import OllamaEmbedding

        embed_model = OllamaEmbedding(

          model_name="nomic-embed-text",

          base_url="http://localhost:11434",

          embed_batch_size=10

        )

        # Chroma with 768-dim vectors for Nomic Embed

        import chromadb

        client = chromadb.Client()

        collection = client.create_collection(

          name="long_docs",

          metadata={"hnsw:space": "cosine"}

          # Nomic Embed vectors: 768 dimensions

        )

Task Prefix Requirement

Nomic Embed Text is trained with task-type prefixes that improve retrieval quality. When embedding documents for indexing, prepend search_document: to the text. When embedding user queries at retrieval time, prepend search_query:. Omitting these prefixes still works but reduces retrieval accuracy by approximately 3–5% on benchmark evaluations. LangChain and LlamaIndex integrations handle this automatically when configured correctly.

Use Cases Where Nomic Embed Excels

Policy and compliance document RAG: HR policies, regulatory filings, compliance manuals, and legal opinions are all document-length texts that benefit from whole-document embedding. Nomic Embed's 8K context handles these without chunking fragmentation.
Technical documentation and runbooks: IT teams embedding long system documentation, API references, and runbooks. Sections stay together as coherent units rather than being split at arbitrary token boundaries.
Research paper and report libraries: Academic and research organizations embedding full paper sections, white papers, and intelligence reports. Long-context embedding preserves the analytical structure of research documents.
Contract and agreement repositories: Legal teams building private contract search tools. Full contract clauses and sections embed as units, making clause-level retrieval accurate without over-chunking.
Knowledge base RAG with large article content: Organizations whose knowledge base articles are long-form — detailed troubleshooting guides, product documentation, training materials — where 256-token chunks lose critical context.
Large corpus deployments with storage constraints: Organizations using Matryoshka truncation to reduce vector storage costs on millions of documents while maintaining acceptable retrieval quality.
OpenWebUI RAG pipeline (simplest path): Nomic Embed Text is the most commonly recommended embedding model for OpenWebUI's built-in document RAG feature. If you are already running OpenWebUI on NoCloudGPT, adding Nomic Embed is a one-click configuration change.

Nomic Embed vs. Other Long-Context Options

Model	Max Tokens	Dimensions	Matryoshka	Open Training	MTEB Score
Nomic Embed v1.5	8,192	768	✓ 64–768	✓ Fully open	62.4
Nomic Embed v2 MOE	8,192	768	✓	✓ Fully open	~64.5
BGE-M3	8,192	1,024	✗	Partial	64.8
All-MiniLM-L6	256	384	✗	Partial	56.3
Mxbai-Embed-Large	512	1,024	✗	Partial	64.7
Snowflake Arctic Embed 2	8,192	1,024	✓	Partial	~65.0

Nomic Embed's unique position: It is the only model in this comparison that offers both 8K token context and fully open training data documentation. For organizations where AI governance requires auditable training provenance AND long-document retrieval capability, Nomic Embed v1.5 or v2 MOE is the only option that satisfies both simultaneously.

Deployment Service Options

Nomic Embed + OpenWebUI

Simplest RAG deployment

Nomic Embed Text v1.5 configured as the embedding model inside OpenWebUI's document RAG feature. Upload documents through the browser interface, query them immediately. No pipeline code required. Ready in under an hour on a Lightsail 4GB instance.

Full Long-Document RAG Stack Most Common

Standard engagement

Nomic Embed v1.5 or v2 MOE with ChromaDB or Qdrant, LangChain or LlamaIndex pipeline, generation model (Llama 3, Mistral, or Qwen), and document ingestion automation. Matryoshka dimension selection included based on corpus analysis.

Governance-Grade Deployment

Compliance-sensitive organizations

Nomic Embed with full training provenance documentation, AI governance package, and audit trail configuration. Includes written documentation of model training lineage suitable for regulatory disclosure and internal AI governance review.

Request Deployment Quote

Related Embedding Models

Need multilingual support with 8K context? BGE-M3 covers 100 languages at similar context length. Need maximum English retrieval quality regardless of context? Mxbai-Embed-Large leads English MTEB. Need the lightest possible footprint? All-MiniLM fits in 90MB of RAM.