Embeddings

Embeddings are the foundation of semantic search and Retrieval-Augmented Generation (RAG).

They transform text, documents, images, and other content into numerical representations that AI systems can search and understand based on meaning rather than exact keywords.

Without embeddings, AI systems cannot effectively retrieve enterprise knowledge.

---

Why Embeddings Matter

Consider the following question:

My VPN access expired. Can you restore it?

The relevant company document may contain:

VPN credentials must be renewed every 90 days and require manager approval.

A traditional keyword search may struggle because:

• “expired” ≠ “renewed”
• “restore” ≠ “approval”
• “VPN access” ≠ “VPN credentials”

Humans understand these concepts are related.

Embeddings allow AI systems to understand this relationship as well.

---

What Is An Embedding?

An embedding is a numerical representation of content.

For example:

"VPN credentials require manager approval"

↓

[0.284, -0.193, 0.729, 0.041, ...]

This list of numbers represents the semantic meaning of the text.

Documents with similar meanings produce vectors that are close together in vector space.

---

Similar Meaning Produces Similar Embeddings

Example:

Text	Semantic Similarity
VPN access expired	High
VPN credentials require renewal	High
Reset employee password	Medium
Company quarterly revenue report	Low

The embedding model learns these relationships automatically.

---

How Embeddings Work In Karyam

When a file is uploaded, Karyam executes the following process:

File Upload
      ↓
Embedding Listener Triggered
      ↓
Document Parsing
      ↓
Chunk Generation
      ↓
Embedding Model
      ↓
Vector Creation
      ↓
Vector Database Storage

The generated embeddings are then available for retrieval by AI agents and workflows.

---

Query Embeddings

Embeddings are generated for both:

Documents

Example:

VPN credentials require approval.

↓

[0.284, -0.193, 0.729, ...]

---

User Questions

Example:

Can you restore my VPN access?

↓

[0.276, -0.201, 0.741, ...]

---

Similarity Search

The vector database compares embeddings and retrieves the closest matches.

User Question
      ↓
Query Embedding
      ↓
Vector Similarity Search
      ↓
Relevant Chunks Retrieved
      ↓
Context Added To Prompt

This process happens automatically during retrieval.

---

Embedding Models

Embedding models are different from chat models.

Their only job is to convert content into vectors.

Examples include:

• Gemini Embeddings
• Voyage AI
• Ollama Embeddings

---

Chat Models vs Embedding Models

Capability	Chat Model	Embedding Model
Generate responses	Yes	No
Hold conversations	Yes	No
Generate embeddings	No	Yes
Semantic search	No	Yes

Both model types are required for Retrieval-Augmented Generation.

---

Supported Embedding Providers

Karyam supports multiple embedding providers.

Examples:

Google Gemini

Examples:

• gemini-embedding-2

---

Voyage AI

Examples:

• voyage-multimodal-3.5

---

Ollama

Examples:

• qwen3-embedding:8b

Ideal for on-premises deployments.

---

Embedding Dimensions

Each embedding model produces vectors of a fixed size.

Examples:

Model	Dimensions
gemini-embedding-2	3072
voyage-multimodal-3.5	1024
qwen3-embedding:8b	4096

Vector databases must support the configured dimensions.

---

Embeddings Are Generated Automatically

In Karyam, users do not manually generate embeddings.

Embeddings are created automatically when:

• Files are uploaded
• Embedding listeners are triggered
• Chunking is completed

This makes knowledge ingestion fully automated.

---

Observability

Embedding operations are fully observable through:

• RAG Runs
• Chunk Metadata
• Embedding Status
• Error Tracking

This allows teams to monitor and debug ingestion pipelines.

---

Example Execution

vpn-policy.pdf uploaded
        ↓
Listener triggered
        ↓
35 chunks generated
        ↓
35 embeddings created
        ↓
Stored in Qdrant
        ↓
RAG Run completed

---

Best Practices

Use A Single Embedding Model Per Vector Database

Changing dimensions requires rebuilding storage.

---

Choose Models Based On Workload

Always choose the models based on the tasks to be performed

---

Monitor Retrieval Quality

Poor retrieval quality is often caused by:

• Weak embedding models
• Poor chunking
• Missing metadata
• Incorrect vector dimensions

---

The Karyam Model

Files
    ↓
Chunking
    ↓
Embeddings
    ↓
Vector Database
    ↓
Retrieval
    ↓
Agents

Embeddings are the bridge between enterprise knowledge and AI reasoning.

---

Next Steps

Continue with:

➡️ File Uploads

Learn how Karyam ingests enterprise content and automatically triggers embedding pipelines.

---