Models
Models are the reasoning engine behind every AI system in Karyam.
Whenever an Agent answers a question, an AI Flow executes, or a document is embedded for retrieval, a Model performs the underlying computation.
Models provide the intelligence that enables AI systems to understand, reason, and generate responses.
Why Models Exist
Section titled “Why Models Exist”Every AI capability depends on a model.
Different models are optimized for different tasks, including:
- Conversational AI
- Reasoning
- Code generation
- Embedding generation
- Vision understanding
- Audio processing
Karyam allows you to configure and manage these models independently from the applications that use them.
Model Types
Section titled “Model Types”Karyam supports two primary model types.
Chat Models
Section titled “Chat Models”Chat models are used by:
- Agents
- AI Flows
- API executions
- Embedded chat widgets
These models perform reasoning, planning, and response generation.
Examples include:
- GPT-4o
- Claude Sonnet
- Gemini
- Bedrock Models
- Ollama Models
Embedding Models
Section titled “Embedding Models”Embedding models convert text into numerical vectors.
These vectors are stored inside Vector Databases and enable semantic search.
Embedding models are used by:
- Embedding Listeners
- Knowledge ingestion
- Retrieval-Augmented Generation (RAG)
Unlike chat models, embedding models do not generate responses.
They generate vector representations of content.
Supported Providers
Section titled “Supported Providers”Karyam supports models from multiple providers.
Examples include:
- ChatGPT
- Google Gemini
- Anthropic
- Perplexity
- AWS Bedrock
- OpenAI Compat
- Cloudflare
- Ollama
- Voyage AI
This allows organizations to choose the models that best meet their performance, cost, and compliance requirements.
Model Configuration
Section titled “Model Configuration”Each model is configured with provider-specific information such as:
- Model name
- Provider
- API credentials
- Endpoint configuration
- Token pricing
- Spending limits
These settings determine how the model behaves within the platform.
Token Pricing
Section titled “Token Pricing”Models can define pricing information for cost tracking.
Typical configuration includes:
| Setting | Description |
|---|---|
| Input Token Cost | Cost per one million input tokens |
| Output Token Cost | Cost per one million output tokens |
| Spending Cap | Maximum allowed spend (0 = unlimited) |
This information powers Karyam’s token usage and cost analytics.
Where Models Are Used
Section titled “Where Models Are Used”Models are shared infrastructure.
A single model may be used by multiple components.
Model ├── Agent ├── AI Flow ├── Embedding Listener └── RAG RetrievalThis allows organizations to centrally manage AI providers while reusing them across multiple AI systems.
Relationship to Other Concepts
Section titled “Relationship to Other Concepts”Models ↓Agents ↓AI Flows ↓Runs
Embedding Models ↓Embedding Listeners ↓Vector Databases ↓RAG RunsModels provide the intelligence layer for both conversational AI and enterprise knowledge retrieval.
The Karyam Philosophy
Section titled “The Karyam Philosophy”AI systems should not be tightly coupled to a single provider.
Karyam separates AI infrastructure from business logic, allowing organizations to evolve their model strategy without rebuilding their AI systems.