Qredence Documentation

Fleet RLM calls models through a stock dspy.LM against an OpenAI-compatible Chat Completions endpoint. You configure the Root and Sub roles separately in config/fleet.toml.

Shipped configuration

The shipped configuration uses a DeepSeek model through the Databricks Unity AI Gateway for both roles.

config/fleet.toml

[llm.root]
model = "uscentral.ai_gateway.deepseek-v4-1-flash-service"
api_key_env = "DATABRICKS_TOKEN"
base_url_env = "FLEET_LLM_BASE_URL"
max_tokens = 16384
timeout_seconds = 300
num_retries = 3
cache = false

[llm.sub]
model = "uscentral.ai_gateway.deepseek-v4-1-flash-service"
api_key_env = "DATABRICKS_TOKEN"
base_url_env = "FLEET_LLM_BASE_URL"
max_tokens = 16384
timeout_seconds = 90
num_retries = 3
temperature = 0
cache = false

Set the two variables in .env:

.env

DATABRICKS_TOKEN=<your-databricks-token>
FLEET_LLM_BASE_URL=https://<workspace-host>/ai-gateway/mlflow/v1

The Databricks DeepSeek endpoint requires an https base URL whose path is exactly /ai-gateway/mlflow/v1.

Use a different provider

Set the model ID

Set model in each role. Fleet adds the openai/ prefix when the ID has no provider prefix, so gpt-4o-mini becomes openai/gpt-4o-mini.

Name the credential variable

Set api_key_env to the name of the environment variable that holds the key. Fleet never reads secrets directly from config/fleet.toml.

Set the endpoint

Set either base_url to a literal URL or base_url_env to a variable name. You can’t set both in one role.

Restart Fleet

Fleet reads config/fleet.toml only at startup.

config/fleet.toml

[llm.root]
model = "openai/gpt-4o"
api_key_env = "OPENAI_API_KEY"
base_url = "https://api.openai.com/v1"

[llm.sub]
model = "openai/gpt-4o-mini"
api_key_env = "OPENAI_API_KEY"
base_url = "https://api.openai.com/v1"
temperature = 0

Role settings

Key	Description
`model`	Model ID sent to the provider.
`api_key_env`	Environment variable that holds the API key.
`base_url`, `base_url_env`	Endpoint URL, or the variable that holds it.
`max_tokens`	Maximum output tokens. Must be at least 1.
`timeout_seconds`	Request timeout. Defaults to 300 for Root and 90 for Sub.
`temperature`	Sampling temperature.
`reasoning_effort`	`none`, `low`, `medium`, or `high`, for models that support it.
`num_retries`	Provider retries. Defaults to 3.
`cache`	Enables the DSPy response cache.

Fleet-RLM

Fleet Prime Agent

Fleet Pi

GEPA Omni

Qredence Skills

Fleet Reasoner

Qredence Plugins

Configure models

Shipped configuration

Use a different provider

Role settings

​Shipped configuration

​Use a different provider

​Role settings

Shipped configuration

Use a different provider

Role settings