Skip to content

model_runtime_configuration.md

Latest commit

 

History

History
92 lines (62 loc) · 2.66 KB

model_runtime_configuration.md

File metadata and controls

92 lines (62 loc) · 2.66 KB
Raw

Model Runtime Configuration

Date reviewed: 2026-06-01

This portfolio runs locally by default. The OpenAI path is optional and intentionally limited to model-facing work:

  • Project 1: grounded answer wording after permission filtering, unsafe-content filtering, citation selection, and abstention logic.
  • Project 2: intent classification before deterministic tool permission and approval checks.

The model is not the security boundary.

Default Optional Model

The default optional model is:

OPENAI_MODEL=gpt-5.2

Reason:

  • OpenAI's current model guide lists gpt-5.2 as the best model for coding and agentic tasks across industries.
  • OpenAI's GPT-5.2 guide recommends the Responses API for reasoning, tool-calling, and multi-turn use cases.
  • The portfolio keeps the local deterministic mode as the verified default so demos do not depend on paid API access.

References:

Environment Variables

$env:OPENAI_API_KEY="..."
$env:OPENAI_MODEL="gpt-5.2"
$env:OPENAI_REASONING_EFFORT="medium"
$env:OPENAI_TEXT_VERBOSITY="low"
$env:COPILOT_MODEL_PROVIDER="openai"
$env:OPS_AGENT_MODEL_ROUTER="openai"

Supported reasoning effort values:

none, low, medium, high, xhigh

Supported verbosity values:

low, medium, high

Project 1 Defaults

Project 1 uses:

OPENAI_REASONING_EFFORT=medium
OPENAI_TEXT_VERBOSITY=low

Rationale:

  • Medium effort is a reasonable default for grounded enterprise answers where precision matters.
  • Low verbosity keeps the generated answer concise because citations and evidence are already shown by application code.
  • Structured output forces the response into an answer/confidence/missing-evidence shape.

Project 2 Defaults

Project 2 uses:

OPENAI_REASONING_EFFORT=low
OPENAI_TEXT_VERBOSITY=low

Rationale:

  • Intent routing should be fast and constrained.
  • The model returns only an enum intent.
  • Approval decisions and side effects remain deterministic application behavior.

Interview Positioning

Use this wording:

I keep local deterministic mode as the verified default, then expose an optional OpenAI Responses API path for the model-facing part of each workflow. I tune model, reasoning effort, verbosity, and structured outputs through environment variables, but I do not move permissions or side-effect authorization into the model. That keeps demos reliable, makes production tuning explicit, and lets evals compare local and model-backed behavior.