Model Providers

The Circuit Breaker Labs CLI supports multiple model providers, allowing you to test any AI model through OpenAI’s API, local Ollama models, or custom endpoints.

Available Providers

The CLI supports three provider types:

OpenAI

OpenAI API and compatible endpoints

Ollama

Local models via Ollama

Custom

Any API via Rhai scripting

Command Structure

Providers are specified as subcommands after the evaluation type:

cbl [top-level-args] <evaluation-type> [evaluation-args] <provider> [provider-args]

Example Breakdown

cbl --output-file result.json \
    single-turn --threshold 0.5 --variations 2 --maximum-iteration-layers 2 \
    openai --model gpt-4o --temperature 1.0

--output-file result.json: Top-level CLI argument
single-turn: Evaluation type
--threshold 0.5 ...: Evaluation arguments
openai: Provider type
--model gpt-4o --temperature 1.0: Provider-specific arguments

OpenAI Provider

The openai provider supports OpenAI’s API and any OpenAI-compatible endpoints.

Basic Usage

export OPENAI_API_KEY="your-api-key"

cbl single-turn \
    --threshold 0.5 \
    --variations 2 \
    --maximum-iteration-layers 2 \
    openai --model gpt-4o

Required Arguments

model

string

required

OpenAI model name (e.g., gpt-4o, gpt-4-turbo, gpt-3.5-turbo) or custom fine-tune ID

api-key

string

required

OpenAI API key. Can be provided via --api-key flag or OPENAI_API_KEY environment variable (recommended)

Optional Arguments

Connection Settings

base-url

string

default:"https://api.openai.com/v1"

OpenAI API base URL for compatible endpoints. Set via --base-url or OPENAI_BASE_URL environment variable.

org-id

string

OpenAI organization ID. Set via --org-id or OPENAI_ORG_ID environment variable.

Sampling Parameters

temperature

float

default:"1.0"

Sampling temperature between 0 and 2. Higher values make output more random.

top-p

float

default:"1.0"

Nucleus sampling parameter. Alternative to temperature.

frequency-penalty

float

default:"0"

Number between -2.0 and 2.0. Penalizes tokens based on frequency in the text so far.

presence-penalty

float

default:"0"

Number between -2.0 and 2.0. Penalizes tokens based on whether they appear in the text so far.

Output Control

max-completion-tokens

integer

Maximum number of tokens to generate in the completion.

stop

string[]

Up to 4 sequences where the API will stop generating. Comma-separated: --stop "\n,END,STOP"

integer

default:"1"

Number of completions to generate per prompt.

Advanced Options

logprobs

boolean

Whether to return log probabilities of output tokens.

top-logprobs

integer

Integer between 0 and 20 specifying number of most likely tokens to return at each position.

logit-bias

map

Modify likelihood of specified tokens. Format: --logit-bias "token_id:bias,token_id:bias"

service-tier

enum

Processing tier: auto, default, flex, scale, or priority

reasoning-effort

enum

Effort for reasoning models: none, minimal, low, medium, high, or xhigh

Example: Testing a Fine-Tune

export OPENAI_API_KEY="your-api-key"
export MY_FINETUNE_ID="ft:gpt-4o-2024-08-06:your-org:your-model:id"

cbl --output-file finetune-results.json \
    single-turn \
    --threshold 0.3 \
    --variations 3 \
    --maximum-iteration-layers 2 \
    openai \
    --model $MY_FINETUNE_ID \
    --temperature 1.2

Ollama Provider

The ollama provider connects to local Ollama models running on your machine or a remote Ollama server.

Basic Usage

Install and Start Ollama

Download from ollama.ai and pull a model:

ollama pull llama3.2

Run Evaluation

cbl single-turn \
    --threshold 0.5 \
    --variations 2 \
    --maximum-iteration-layers 2 \
    ollama --model llama3.2

cbl single-turn \
    --threshold 0.5 \
    --variations 2 \
    --maximum-iteration-layers 2 \
    ollama --model llama3.2

Required Arguments

model

string

required

Ollama model name (e.g., llama3.2, mistral, phi3)

Optional Arguments

Connection Settings

base-url

string

default:"http://localhost:11434"

Ollama server URL. Set via --base-url or OLLAMA_BASE_URL environment variable.

Sampling Parameters

temperature

float

default:"0.8"

Model temperature - higher values make answers more creative.

top-k

integer

default:"40"

Reduces probability of generating nonsense. Higher = more diverse.

top-p

float

default:"0.9"

Works with top-k. Higher values lead to more diverse text.

repeat-penalty

float

default:"1.1"

How strongly to penalize repetitions.

Generation Control

num-predict

integer

default:"128"

Maximum tokens to predict. Use -1 for infinite, -2 to fill context.

stop

string[]

Stop sequences. Can be specified multiple times.

seed

integer

default:"0"

Random number seed for reproducible generation.

Performance Tuning

num-ctx

integer

default:"2048"

Size of the context window.

num-gpu

integer

Number of layers to send to GPU(s).

num-thread

integer

Number of threads to use during computation.

Advanced Sampling

mirostat

integer

default:"0"

Enable Mirostat sampling: 0=disabled, 1=Mirostat, 2=Mirostat 2.0

mirostat-eta

float

default:"0.1"

Mirostat learning rate.

mirostat-tau

float

default:"5.0"

Mirostat tau - controls balance between coherence and diversity.

tfs-z

float

default:"1"

Tail free sampling - reduces impact of less probable tokens.

Custom Provider

The custom provider allows you to integrate any API using Rhai scripting. This is perfect for:

Proprietary model APIs
Custom inference endpoints
Non-OpenAI-compatible services
Internal model deployments

For detailed information on creating custom providers, see the Custom Providers guide.

Basic Usage

cbl single-turn \
    --threshold 0.5 \
    --variations 2 \
    --maximum-iteration-layers 2 \
    custom \
    --url https://your-api.com/completions \
    --script ./providers/your-provider.rhai

Required Arguments

url

string

required

The endpoint URL to POST requests to

script

path

required

Path to the Rhai script file that translates requests/responses

Authentication

Custom providers support authentication via request headers:

export CUSTOM_API_KEY="your-api-key"

cbl single-turn \
    --threshold 0.5 \
    --variations 2 \
    --maximum-iteration-layers 2 \
    custom \
    --url https://your-api.com/completions \
    --script ./provider.rhai

See the Custom Providers guide for complete examples with authentication and complex request schemas.

Provider Comparison

Feature	OpenAI	Ollama	Custom
Setup Difficulty	Easy	Medium	Advanced
Cost	Pay per token	Free (local compute)	Varies
Latency	Low (cloud)	Very low (local)	Varies
Model Selection	OpenAI models + fine-tunes	Open-source models	Any model
Configuration	Built-in parameters	Built-in parameters	Script-based
Best For	Production testing, OpenAI models	Local development, open-source models	Custom APIs, proprietary models
Authentication	API key	None (local)	Custom via headers
Offline Usage	No	Yes	Depends

Environment Variables

Circuit Breaker Labs

CBL_API_KEY

string

required

Your Circuit Breaker Labs API key. Required for all evaluations.

export CBL_API_KEY="your-cbl-api-key"

OpenAI Provider

OPENAI_API_KEY

string

required

OpenAI API key

OPENAI_BASE_URL

string

Custom base URL for OpenAI-compatible endpoints

OPENAI_ORG_ID

string

OpenAI organization ID

Ollama Provider

OLLAMA_BASE_URL

string

default:"http://localhost:11434"

Ollama server URL

Custom Provider

Custom providers can read any environment variables your Rhai script accesses. Common patterns:

export CUSTOM_API_KEY="your-key"
export CUSTOM_API_URL="https://your-api.com"

Common Scenarios

Testing Multiple Models

# Test OpenAI's GPT-4
cbl single-turn --threshold 0.5 --variations 2 --maximum-iteration-layers 2 \
    openai --model gpt-4o

# Test local Llama model
cbl single-turn --threshold 0.5 --variations 2 --maximum-iteration-layers 2 \
    ollama --model llama3.2

# Compare results to find the safest model for your use case

CI/CD Integration

# GitHub Actions example
- name: Run safety evaluation
  env:
    CBL_API_KEY: ${{ secrets.CBL_API_KEY }}
    OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
  run: |
    cbl --output-file results.json \
      single-turn --threshold 0.6 --variations 2 --maximum-iteration-layers 2 \
      openai --model gpt-4o

Development Workflow

# 1. Quick local testing with Ollama
ollama pull llama3.2
cbl single-turn --threshold 0.5 --variations 1 --maximum-iteration-layers 1 \
    ollama --model llama3.2

# 2. Comprehensive testing with OpenAI before deployment
cbl single-turn --threshold 0.6 --variations 3 --maximum-iteration-layers 2 \
    openai --model gpt-4o

# 3. Multi-turn testing for production validation
cbl multi-turn --threshold 0.6 --max-turns 8 \
    --test-types user_persona,semantic_chunks \
    openai --model gpt-4o

Troubleshooting

OpenAI: 'Invalid API Key' Error

Verify your API key is correct: echo $OPENAI_API_KEY
Ensure no extra spaces or quotes in the environment variable
Check that your API key has sufficient credits

Ollama: 'Connection Refused' Error

Verify Ollama is running: ollama list
Check the base URL is correct
Ensure the model is pulled: ollama pull llama3.2

Custom: 'Script Error' Messages

Verify your Rhai script syntax is correct
Check that build_request and parse_response functions exist
Test your script with simple inputs first
See Custom Providers guide for debugging tips

Next Steps

Custom Providers

Learn how to create custom providers with Rhai scripting

Single-Turn Evaluations

Configure and run single-turn safety tests

Get Started

Guides

Advanced

Available Providers

OpenAI

Ollama

Custom

Command Structure

OpenAI Provider

Basic Usage

Required Arguments

Optional Arguments

Example: Testing a Fine-Tune

Ollama Provider

Basic Usage

Required Arguments

Optional Arguments

Custom Provider

Basic Usage

Required Arguments

Authentication

Provider Comparison

Environment Variables

Common Scenarios

Troubleshooting

Next Steps

Custom Providers

Single-Turn Evaluations

Get Started

Guides

Advanced

​Available Providers

OpenAI

Ollama

Custom

​Command Structure

​OpenAI Provider

​Basic Usage

​Required Arguments

​Optional Arguments

​Example: Testing a Fine-Tune

​Ollama Provider

​Basic Usage

​Required Arguments

​Optional Arguments

​Custom Provider

​Basic Usage

​Required Arguments

​Authentication

​Provider Comparison

​Environment Variables

​Common Scenarios

​Troubleshooting

​Next Steps

Custom Providers

Single-Turn Evaluations

Available Providers

Command Structure

OpenAI Provider

Basic Usage

Required Arguments

Optional Arguments

Example: Testing a Fine-Tune

Ollama Provider

Basic Usage

Required Arguments

Optional Arguments

Custom Provider

Basic Usage

Required Arguments

Authentication

Provider Comparison

Environment Variables

Common Scenarios

Troubleshooting

Next Steps