Vector Databases

Vector databases are the foundation of Retrieval-Augmented Generation (RAG) in Karyam.

They store numerical representations of enterprise knowledge, allowing AI systems to retrieve information based on meaning rather than exact keywords.

Without vector databases, AI systems cannot efficiently search and reason over organizational knowledge.

Why Traditional Databases Are Not Enough

Traditional databases excel at exact matching.

For example:

SELECT * FROM documents
WHERE title = 'VPN Policy'

This works well when you know exactly what you’re looking for.

However, employees rarely ask questions this way.

Instead they ask:

My VPN access expired. Can you restore it?

The document may never contain the phrase:

VPN access expired

Instead it may contain:

VPN credentials must be renewed every 90 days.

Keyword search struggles here.

Semantic search understands that both statements refer to the same concept.

What Is a Vector?

Embedding models convert text into numerical representations called vectors.

Example:

"VPN credentials expire after 90 days"

↓

[0.12, -0.45, 0.81, 0.03, ...]

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

This allows AI systems to search by meaning rather than exact wording.

How Vector Search Works

Employee Question
        ↓
Embedding Model
        ↓
Query Vector
        ↓
Vector Similarity Search
        ↓
Relevant Chunks Retrieved
        ↓
Context Sent To Agent

The AI model never searches documents directly.

Instead it searches embeddings stored in the vector database.

Karyam Architecture

In Karyam, vector databases are one component of the retrieval pipeline:

Files
    ↓
Embedding Listener
    ↓
Chunking
    ↓
Embedding Model
    ↓
Vector Database
    ↓
Retrieval
    ↓
Agent

This architecture provides:

Flexible storage providers
Multiple embedding models
Event-driven ingestion
Retrieval observability
Enterprise scalability

Supported Providers

Karyam currently supports:

Qdrant

Purpose-built vector database optimized for semantic search workloads.

Best for:

Dedicated RAG infrastructure
Large-scale deployments
High retrieval throughput
Multi-tenant environments

Advantages:

Native vector indexing
High performance similarity search
Rich filtering capabilities
Horizontal scalability

PGVector

PostgreSQL extension that adds vector search capabilities.

Best for:

Existing PostgreSQL environments
Smaller deployments
Simpler infrastructure management
On-premises installations

Advantages:

Reuse existing PostgreSQL infrastructure
SQL-based filtering
Simplified operations
Lower infrastructure complexity

Choosing a Provider

Requirement	Recommended Provider
Maximum performance	Qdrant
Existing PostgreSQL environment	PGVector
Large enterprise deployment	Qdrant
Small team deployment	PGVector
Simplified operations	PGVector
High-scale retrieval	Qdrant

Storage Model in Karyam

Each vector database contains:

Embeddings
Chunk metadata
Source file references
Workspace isolation information

Example:

Vector Database
│
├── Chunk
│   ├── Content
│   ├── Embedding
│   ├── File Path
│   ├── Metadata
│   └── Workspace UUID
│
├── Chunk
│
└── Chunk

This enables:

Multi-tenant environments
File-level traceability
Source attribution
Retrieval debugging

Vector Dimensions

The size of an embedding vector depends on the embedding model used.

Examples:

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

Vector databases must support the embedding dimensions produced by the configured embedding model.

Metadata Filtering

Karyam supports metadata-aware retrieval.

Examples:

Only search:

- HR documents
- Finance policies
- Workspace A
- Files uploaded after July 2026

This improves retrieval accuracy and reduces irrelevant context.

Observability

Every interaction with the vector database is observable through:

RAG Runs
Retrieval Logs
Chunk Metadata
Source Attribution

This makes semantic retrieval debuggable in production environments.

Example Retrieval Flow

Employee:
"My VPN access expired."

Agent:
↓
Generate Query Embedding
↓
Search Vector Database
↓
Retrieve VPN Policy Chunks
↓
Generate Response

Retrieved context:

VPN credentials expire every 90 days and require manager approval for renewal.

The agent can now answer using company policy rather than model assumptions.

Best Practices

Use Dedicated Databases For Large Domains

Example:

HR Database
Finance Database
Engineering Database
Legal Database

Keep Embedding Dimensions Consistent

Changing embedding dimensions requires rebuilding vector storage.

Choose embedding models carefully.

Use Metadata

Metadata improves retrieval quality and observability.

Always preserve:

File paths
Categories
Workspace identifiers
Source information

Monitor Retrieval Quality

Poor retrieval quality usually originates from:

Incorrect chunking
Weak embedding models
Missing metadata
Oversized documents

RAG Runs help identify these issues quickly.

Next Steps

Now that you understand vector storage, continue with:

➡️ Embeddings

Learn how embedding models convert enterprise knowledge into vectors that can be searched semantically.