Configuration

Overview

Configuration controls the LLM backend used by the review path. The deterministic scoring path (used for debugging, testing, and illustration) requires no configuration. Settings can be provided as a YAML file, a Python dict, or environment variables.

YAML configuration

Create a config file and pass the path to evaluate_confidence() or Evaluate():

from impact_engine_evaluate import evaluate_confidence

result = evaluate_confidence("review_config.yaml", "path/to/job-dir/")

Anthropic:

backend:
  model: claude-sonnet-4-6
  temperature: 0.0
  max_tokens: 4096

Ollama (local):

backend:
  model: ollama_chat/llama3.2   # routes to http://localhost:11434
  temperature: 0.0
  max_tokens: 2048
  # api_base: "http://my-ollama-server:11434"  # custom endpoint

The ollama_chat/<model> prefix is routed to http://localhost:11434 by litellm automatically. Any extra keys (e.g. api_base) are forwarded as kwargs to litellm.completion().

Dict configuration

from impact_engine_evaluate import evaluate_confidence

result = evaluate_confidence(
    {"backend": {"model": "gpt-4o", "temperature": 0.0, "max_tokens": 4096}},
    "path/to/job-dir/",
)

Environment variables

Environment variables override any values from YAML or dict sources. Pass config=None (the default) to use environment variables alone.

Variable	Description	Default
`REVIEW_BACKEND_MODEL`	Model identifier (any LiteLLM-supported model)	`claude-sonnet-4-5-20250929`
`REVIEW_BACKEND_TEMPERATURE`	Sampling temperature	`0.0`
`REVIEW_BACKEND_MAX_TOKENS`	Maximum tokens per completion	`4096`

export REVIEW_BACKEND_MODEL=gpt-4o

Backend parameter reference

Parameter	Type	Description
`model`	str	Model identifier passed to `litellm.completion()`. Any model supported by LiteLLM.
`temperature`	float	Sampling temperature. `0.0` produces deterministic output.
`max_tokens`	int	Maximum tokens in the LLM response.

Additional keys are forwarded as keyword arguments to litellm.completion() via the extra dict.

Dependencies

All review dependencies are core requirements (installed automatically):

Package	Role
LiteLLM	100+ LLM providers via unified API
Jinja2	Prompt template rendering
PyYAML	YAML config and prompt loading

pip install impact-engine-evaluate

Precedence

When the same parameter appears in multiple sources, the resolution order is:

Environment variables (highest priority)
YAML file or dict values
Built-in defaults (lowest priority)