Overview Fenic supports multiple LLM providers with provider-specific configuration options. Each model type has its own configuration class with rate limiting and provider-specific parameters.
Language Models OpenAILanguageModel Configuration for OpenAI language models including GPT-4 and o-series reasoning models.
from fenic.api.session.config import OpenAILanguageModel model = OpenAILanguageModel( model_name = "gpt-4.1-nano" , rpm = 100 , tpm = 1000 , profiles = None , default_profile = None )
model_name
OpenAILanguageModelName
required
The name of the OpenAI model to use. Examples: "gpt-4.1-nano", "o4-mini", "gpt-5.1"
Requests per minute limit. Must be greater than 0.
Tokens per minute limit. Must be greater than 0.
Optional mapping of profile names to profile configurations. Allows the same model to be used with different settings. Show OpenAI Profile Options
Reasoning effort level for gpt5 and o-series models. Valid values: "none", "minimal", "low", "medium", "high"
gpt-5.1 models : defaults to "none" (disabled reasoning), does NOT support "minimal"gpt-5 models : defaults to "minimal", does NOT support "none"o-series models : defaults to "low", does NOT support "none" or "minimal" Verbosity level for gpt5/gpt5.1 models only.
The name of the default profile to use if profiles are configured.
Example with Profiles: model = OpenAILanguageModel( model_name = "o4-mini" , rpm = 100 , tpm = 1000 , profiles = { "fast" : OpenAILanguageModel.Profile( reasoning_effort = "low" ), "thorough" : OpenAILanguageModel.Profile( reasoning_effort = "high" ), }, default_profile = "fast" , )
AnthropicLanguageModel Configuration for Anthropic Claude models with separate input/output token limits.
from fenic.api.session.config import AnthropicLanguageModel model = AnthropicLanguageModel( model_name = "claude-3-5-haiku-latest" , rpm = 100 , input_tpm = 10000 , output_tpm = 4000 , profiles = None , default_profile = None )
model_name
AnthropicLanguageModelName
required
The name of the Anthropic model to use. Examples: "claude-3-5-haiku-latest", "claude-opus-4-0"
Requests per minute limit. Must be greater than 0.
Input tokens per minute limit. Must be greater than 0.
Output tokens per minute limit. Must be greater than 0.
Optional mapping of profile names to profile configurations. Show Anthropic Profile Options
Thinking budget in tokens. If not provided, thinking will be disabled for the profile. Minimum supported by Anthropic: 1024 tokens If thinking_token_budget is set, temperature cannot be customized.
The name of the default profile to use if profiles are configured.
Example with Thinking Profiles: model = AnthropicLanguageModel( model_name = "claude-opus-4-0" , rpm = 100 , input_tpm = 10000 , output_tpm = 4000 , profiles = { "thinking_disabled" : AnthropicLanguageModel.Profile(), "fast" : AnthropicLanguageModel.Profile( thinking_token_budget = 1024 ), "thorough" : AnthropicLanguageModel.Profile( thinking_token_budget = 4096 ), }, default_profile = "fast" , )
GoogleDeveloperLanguageModel Configuration for Gemini models accessible through Google Developer AI Studio (requires GOOGLE_API_KEY).
from fenic.api.session.config import GoogleDeveloperLanguageModel model = GoogleDeveloperLanguageModel( model_name = "gemini-2.0-flash" , rpm = 100 , tpm = 10000 , profiles = None , default_profile = None )
model_name
GoogleDeveloperLanguageModelName
required
The name of the Google Developer model to use. Examples: "gemini-2.0-flash", "gemini-2.5-flash", "gemini-3.5-flash"
Requests per minute limit. Must be greater than 0.
Tokens per minute limit. Must be greater than 0.
Optional mapping of profile names to profile configurations. Show Google Profile Options
For gemini-2.5 and earlier models only. Thinking budget in tokens.
If not provided or set to 0, thinking is disabled (not supported on gemini-2.5-pro)
Set to -1 for automatic budget based on prompt complexity
Range: -1 to 32767
Gemini models treat this as a suggestion, not a hard limit. Models may generate more thinking tokens than the budget.
For gemini-3+ models only. Set the thinking level.Options: "high", "medium", "low", "minimal" Mutually exclusive with thinking_token_budget.
For gemini-3+ models only. Media resolution for PDF processing.Options: "low", "medium", "high", "ultra_high" Affects token cost per page.
The name of the default profile to use if profiles are configured.
Example with Thinking Profiles (gemini-2.5): model = GoogleDeveloperLanguageModel( model_name = "gemini-2.5-flash" , rpm = 100 , tpm = 10000 , profiles = { "thinking_disabled" : GoogleDeveloperLanguageModel.Profile(), "fast" : GoogleDeveloperLanguageModel.Profile( thinking_token_budget = 1024 ), "auto" : GoogleDeveloperLanguageModel.Profile( thinking_token_budget =- 1 ), "thorough" : GoogleDeveloperLanguageModel.Profile( thinking_token_budget = 8192 ), }, default_profile = "fast" , )
Example with Thinking Levels (gemini-3+): model = GoogleDeveloperLanguageModel( model_name = "gemini-3.5-flash" , rpm = 100 , tpm = 10000 , profiles = { "minimal" : GoogleDeveloperLanguageModel.Profile( thinking_level = "minimal" ), "low" : GoogleDeveloperLanguageModel.Profile( thinking_level = "low" ), "high" : GoogleDeveloperLanguageModel.Profile( thinking_level = "high" ), }, default_profile = "low" , )
GoogleVertexLanguageModel Configuration for Gemini models accessible through Google Vertex AI (requires Google Cloud credentials).
Has the same parameters and profile options as GoogleDeveloperLanguageModel.
from fenic.api.session.config import GoogleVertexLanguageModel model = GoogleVertexLanguageModel( model_name = "gemini-2.0-flash" , rpm = 100 , tpm = 10000 )
OpenRouterLanguageModel Configuration for OpenRouter language models with advanced provider routing (requires OPENROUTER_API_KEY).
from fenic.api.session.config import OpenRouterLanguageModel model = OpenRouterLanguageModel( model_name = "anthropic/claude-3-5-sonnet" , profiles = { ... }, default_profile = "default" , structured_output_strategy = None )
Model identifier in {provider}/{model} format. Example: "anthropic/claude-3-5-sonnet"
Optional mapping of profile names to profile configurations. Show OpenRouter Profile Options
reasoning_effort
Literal['high', 'medium', 'low']
OpenAI-style reasoning effort. If the model supports reasoning but not reasoning_effort, a reasoning_max_tokens will be calculated.
Token budget for reasoning (for Anthropic, Gemini, etc.). If the model supports reasoning but not reasoning_max_tokens, a reasoning_effort will be calculated.
List of fallback models to use if the primary model is unavailable. Maximum 3 models.
Parsing engine for PDF files. Options: "native", "pdf-text", "mistral-ocr" By default, the model’s native parsing engine is used. Note: "mistral-ocr" incurs additional costs.
Provider routing preferences. Show Provider Routing Options
List of providers to try in order. Example: ["Anthropic", "Amazon Bedrock"]
Provider routing preference:
"price" - Route to cheapest provider first"throughput" - Route to highest throughput provider"latency" - Route to lowest latency provider Allowed quantizations. Example: ["bf16", "fp16"] Note: Many providers report "unknown" quantization.
Data collection preference:
"allow" - Allow providers that may store/train on prompts"deny" - Only use providers that don’t collect data Only include these providers when routing. Example: ["Anthropic"]
Exclude these providers when routing.
Maximum prompt price ($USD per 1M tokens).
Maximum completion price ($USD per 1M tokens).
The name of the default profile to use if profiles are configured.
structured_output_strategy
Strategy for structured output if model supports both tool calling and structured outputs:
"prefer_tools" - Prefer using tools over response format"prefer_response_format" - Prefer using response format over tools Example with Provider Routing: model = OpenRouterLanguageModel( model_name = "anthropic/claude-sonnet-4-0-latest" , profiles = { "default" : OpenRouterLanguageModel.Profile( provider = OpenRouterLanguageModel.Provider( only = [ "Anthropic" ], # Only route to Anthropic, not AWS/GCP sort = "price" ) ) }, )
Example with Throughput Optimization: model = OpenRouterLanguageModel( model_name = "qwen/qwen3-next-80b-a3b-instruct" , profiles = { "default" : OpenRouterLanguageModel.Profile( provider = OpenRouterLanguageModel.Provider( sort = "throughput" , data_collection = "deny" , quantizations = [ "bf16" ] # Only bf16, no fp8 ) ) }, )
Embedding Models OpenAIEmbeddingModel Configuration for OpenAI embedding models.
from fenic.api.session.config import OpenAIEmbeddingModel model = OpenAIEmbeddingModel( model_name = "text-embedding-3-small" , rpm = 100 , tpm = 1000 )
model_name
OpenAIEmbeddingModelName
required
The name of the OpenAI embedding model. Examples: "text-embedding-3-small", "text-embedding-3-large"
Requests per minute limit. Must be greater than 0.
Tokens per minute limit. Must be greater than 0.
GoogleDeveloperEmbeddingModel Configuration for Google embedding models via Developer AI Studio.
from fenic.api.session.config import GoogleDeveloperEmbeddingModel model = GoogleDeveloperEmbeddingModel( model_name = "gemini-embedding-001" , rpm = 100 , tpm = 10000 , profiles = None , default_profile = None )
model_name
GoogleDeveloperEmbeddingModelName
required
The name of the Google Developer embedding model.
Requests per minute limit. Must be greater than 0.
Tokens per minute limit. Must be greater than 0.
Optional mapping of profile names to profile configurations. Dimensionality of the embedding output. Range: 1 to 3072. If not provided, uses the model’s default dimensionality.
task_type
GoogleEmbeddingTaskType
default: "SEMANTIC_SIMILARITY"
Type of task for the embedding model. Options:
"SEMANTIC_SIMILARITY""CLASSIFICATION""CLUSTERING""RETRIEVAL_DOCUMENT""RETRIEVAL_QUERY""CODE_RETRIEVAL_QUERY""QUESTION_ANSWERING""FACT_VERIFICATION" The name of the default profile to use if profiles are configured.
Example with Profiles: model = GoogleDeveloperEmbeddingModel( model_name = "gemini-embedding-001" , rpm = 100 , tpm = 10000 , profiles = { "default" : GoogleDeveloperEmbeddingModel.Profile(), "high_dim" : GoogleDeveloperEmbeddingModel.Profile( output_dimensionality = 3072 ), "retrieval" : GoogleDeveloperEmbeddingModel.Profile( output_dimensionality = 1536 , task_type = "RETRIEVAL_DOCUMENT" ), }, default_profile = "default" , )
GoogleVertexEmbeddingModel Configuration for Google embedding models via Vertex AI.
Has the same parameters and profile options as GoogleDeveloperEmbeddingModel, but with a minimum output_dimensionality of 768.
from fenic.api.session.config import GoogleVertexEmbeddingModel model = GoogleVertexEmbeddingModel( model_name = "gemini-embedding-001" , rpm = 100 , tpm = 10000 )
CohereEmbeddingModel Configuration for Cohere embedding models.
from fenic.api.session.config import CohereEmbeddingModel model = CohereEmbeddingModel( model_name = "embed-v4.0" , rpm = 100 , tpm = 50000 , profiles = None , default_profile = None )
model_name
CohereEmbeddingModelName
required
The name of the Cohere embedding model.
Requests per minute limit. Must be greater than 0.
Tokens per minute limit. Must be greater than 0.
Optional mapping of profile names to profile configurations. Dimensionality of the embedding output. Range: 1 to 1536. If not provided, uses the model’s default dimensionality.
input_type
CohereEmbeddingTaskType
default: "search_document"
Type of input text. Options:
"search_document""search_query""classification""clustering" The name of the default profile to use if profiles are configured.
Example with Profiles: model = CohereEmbeddingModel( model_name = "embed-v4.0" , rpm = 100 , tpm = 50000 , profiles = { "high_dim" : CohereEmbeddingModel.Profile( output_dimensionality = 1536 , input_type = "search_document" ), "classification" : CohereEmbeddingModel.Profile( output_dimensionality = 1024 , input_type = "classification" ), }, default_profile = "high_dim" , )
Environment Variables Each model provider requires specific environment variables:
OpenAI : OPENAI_API_KEYAnthropic : ANTHROPIC_API_KEYGoogle Developer : GOOGLE_API_KEYGoogle Vertex : Google Cloud credentials (ADC or service account)OpenRouter : OPENROUTER_API_KEYCohere : COHERE_API_KEY
