Skip to content

Providers providers

prich supports multiple provider backends. You can set a settings.default_provider, assign provider to a specific template using settings.provider_assignments (see Template Provider Assignments), and override it per run with --provider <provider-name>.

Currently prich supports OpenAI, Ollama, and MLX LM providers, plus STDIN Consumer and simple echo output.

Echo echo

Render the prompt but do not call a model (good for debugging, CI dry-runs, copy-paste into other tool, or pipe to stdin).

show_prompt:
  provider_type: "echo"
  mode: "flat"      # see "Modes"

OpenAI-compatible HTTP openai

Generic HTTP client for OpenAI and compatible services.

openai-gpt4o:
  provider_type: "openai"
  configuration:                    # would be sent as OpenAI(**configuration)
    api_key: "${OPENAI_API_KEY}"
    base_url: "https://api.openai.com/v1"   # <— use a real endpoint
  options:                          # would be sent as client.chat.completions.create(**options)
    model: "gpt-4o"

Ollama HTTP (using /generate endpoint) ollama

Local server (using http://localhost:11434 by default)

qwen3-8b:
  provider_type: "ollama"
  model: "llama3.1:8b"
  stream: true
  options:
    num_predict: 3000

Advanced/raw prompt control: 

qwen3-8b-custom:
  provider_type: "ollama"
  model: "qwen3:8b"    # model name installed in your Ollama
  mode: "flat"         # optional [str] name of prich config provider prompt template mode (see config.yaml `provider_modes`)
  base_url: "http://localhost:11434"  # optional [str] - Ollama server URL 
  think: true          # optional [bool] - for thinking models only to enable/disable
  stream: true         # optional [bool] - stream output
  suffix: ""           # optional [str]
  raw: true            # optional [bool] - send prompt without default model prompt template, use with `mode: "..."`
  format: "json"       # optional [dict | str] - "json" or json schema for specific output format
  options:             # optional [dict] - model options (vary depending on used model)
    num_predict: 5000

STDIN consumer (bridge to CLIs) stdin_consumer

Send prompts to a command via STDIN and read STDOUT (e.g., q chat).

qchat:
  provider_type: "stdin_consumer"
  mode: "flat"
  call: "q"                      # executable file
  args: ["chat", "--no-interactive"]  # optional arguments
  strip_output_prefix: "> "      # optional [str] - strip prefix text from step output
  slice_output_start: 0          # optional [int] - slice step output text starting from N
  slice_output_end: -1           # optional [int] - slice step output text ending at N (negative counts from the back)