Models

Raikoo uses a layered model system to connect AI services to workflows, agents, and chat sessions. Understanding this hierarchy helps you choose the right model for every context and ensure your organization's AI capabilities are configured correctly.

AI Providers

An AI provider is an organization-level connection to an external LLM service. Each provider holds the API credentials and configuration needed to access that service, and exposes a set of models to agents, workflows, and chat sessions within the organization.

Organizations can configure multiple providers, including multiple providers of the same type (for example, separate OpenAI accounts for development and production).

Supported Provider Types

Provider Type	Description
OpenAI	OpenAI's API (GPT, o-series, DALL-E, embeddings)
Anthropic	Anthropic's Claude model family
Google Generative AI	Google's Gemini model family
Azure OpenAI	OpenAI models hosted on Azure infrastructure
Ollama	Locally-hosted open-source models
OpenAI-Compatible	Any API that follows the OpenAI API format
Amazon Bedrock	AWS-hosted foundation models
Groq	High-speed inference for open-source models
Mistral	Mistral AI model family
Cohere	Cohere's language and embedding models
Perplexity	Perplexity's search-augmented models
DeepSeek	DeepSeek's reasoning and coding models
OpenRouter	Multi-provider routing service

For setup instructions, see the AI Providers user guide.

Models

A model is an individual LLM available through a provider. Each model has an ID (as returned by the provider's API) and optional capability metadata that Raikoo uses to make intelligent decisions about how to interact with it.

Model Capabilities

Raikoo tracks the following capability information for each model:

Capability	Description
Context window	Maximum number of tokens the model can process in a single request (input and output combined)
Max output tokens	Maximum number of tokens the model can generate in a single response
Input modalities	Types of content the model can receive: text, image, audio, video
Output modalities	Types of content the model can produce: text, image, audio, video, embeddings
System prompt	Whether the model supports a system or developer prompt for guiding behavior
Tools	Whether the model supports function calling and tool use
Reasoning effort	Supported levels for extended thinking (e.g., low, medium, high)
Max budget tokens	Maximum token budget that can be allocated for reasoning

Raikoo uses this data to adapt the user interface and runtime behavior. For example, image attachment controls are only shown for models that support vision input, and reasoning effort controls appear only for models that support extended thinking.

Model Resolution

When a provider returns a list of model IDs, Raikoo runs each one through a resolution pipeline to find matching capability data from its built-in library of known models:

Exact match -- The model ID directly matches a known definition (e.g., gpt-4o matches the GPT-4o definition).
Inferred match -- The model ID contains a known model name as a stem (e.g., a deployment named gpt-4o-2024-08-06 or company-gpt-4o-prod infers capabilities from the gpt-4o definition).
Unknown -- No match is found. The model has no pre-loaded capability data and may need manual configuration.

Resolution happens automatically during model discovery. Administrators can review and override the resolved capabilities for any model.

Capability Confidence

Separate from resolution confidence, capability confidence indicates how complete a model's capability data is:

Level	Meaning
Full info	Context window, capabilities, and modalities are all defined
Partial info	Some capability fields are defined, but others are missing
No info	Only the model ID is known; no capability data is available

When capability data is missing, features that depend on it (such as vision attachment gating, reasoning controls, or context window warnings) may not function correctly. Administrators can fill in missing capability data through the model editor in the AI Providers settings.

Model Families

A model family is a named logical grouping of models with an ordered fallback sequence. When a workflow or agent targets a family rather than a specific model, Raikoo selects the best available model from the group at runtime. If the preferred model is unavailable, the next model in the family is tried automatically.

Model families provide several advantages:

Abstraction -- Workflows and agents can target a capability level (e.g., "fast" or "capable") without hard-coding a specific model or provider.
Resilience -- If a model or provider becomes unavailable, the family automatically falls back to the next option.
Consistency -- Teams standardize on a family rather than a specific model version, making upgrades seamless.

Model families are defined at the organization level and can include models from multiple providers.

Model Selection

Models can be selected at several levels in Raikoo. More specific selections override broader defaults:

Organization Default

The organization can configure a default model or model family used when no other selection is specified. This serves as the baseline for all AI operations.

Agent Configuration

Each agent has a fallback model setting -- the model used when the agent runs without a more specific override. This can be a specific model from a provider or a model family.

Workflow Operations

Individual operations within a workflow can specify a model selector, overriding the agent's default model for that specific step. This allows a single workflow to use different models for different tasks (e.g., a fast model for classification and a capable model for generation).

Chat Sessions

In the Chat tool, users can select a specific model and provider or a model family from the settings panel. This selection is preserved when saving the chat session.

Selection Priority

When Raikoo needs to determine which model to use, it follows this priority order (most specific wins):

Chat/operation override -- A model explicitly selected for this specific chat session or workflow operation
Agent fallback model -- The model configured on the agent
Organization default -- The organization-level default model or family

Next Steps

Set up AI providers and manage models: AI Providers
Configure agents with specific models: Agents
Use models in chat sessions: Chat