Tuning

Configuration

QUASAR is primarily configured via environment variables, with the same controls exposed in the browser Settings panel and CLI settings view.

Core Environment Variables

Variable	Required	Purpose	Default
`MODEL`	Yes	Base model name for the system.	None
`MODEL_API_KEY`	Yes	API key for the base model provider.	None
`OPENAI_API_BASE`	No	Base URL for OpenAI-compatible global endpoints (same effect as `API_BASE_URL`).	None
`API_BASE_URL`	No	Alias for `OPENAI_API_BASE` when routing the primary model through an OpenAI-compatible HTTP API.	None
`ACCURACY`	No	Planning/execution rigor: `eco`, `standard`, `pro`, or `adaptive` (experimental: scales numerical rigor with workflow stage).	`standard`
`GRANULARITY`	No	Task decomposition depth: `low`, `medium`, `high`, or `adaptive` (experimental: scales task count with perceived complexity).	`adaptive`
`AUTO_IMPROVE_CYCLES`	No	Number of automatic auto-improve follow-up runs after a successful user-started run.	`0`
`PMG_MAPI_KEY`	No	Materials Project database access	None
`IF_RESTART`	No	Resume from checkpoint when present.	`false`

Advanced Environment Variables

These settings are helpful for specialized workflows and can usually be left at their defaults unless you have a specific reason to change them.

Variable	Required	Purpose	Default
`CONTEXT_THRESHOLD`	No	Context compression trigger level: `low` = 20%, `medium` = 40%, `hard` = 60% of model context.	`medium`
`ENABLE_RAG`	No	Enable documentation retrieval.	`true`
`NUM_CORES`	No	Override physical core detection.	`Auto`
`AUTO_CONFIRM_PLAN`	No	If `true`/`1`/`yes`/`on`, skips interactive plan confirmation (useful for headless or batch runs).	unset (confirmation on)

Long-running Python check-ins are scheduled by the Operator per tool call through execute_python(check_in_after=...) and can be rescheduled with continue_execution(next_check_in_after=...); there is no active CHECK_INTERVAL environment variable.

When resuming from a checkpoint, ACCURACY and GRANULARITY are locked to the saved run values so an interrupted workflow continues with the same planning assumptions.

OpenAI-compatible LLM endpoints

QUASAR routes requests through an OpenAI-compatible HTTP client when:

OPENAI_API_BASE or API_BASE_URL is set — all traffic for the primary model uses that base URL (for example a gateway, proxy, or third-party OpenAI-compatible host), regardless of model name.
The model name looks like OpenAI gpt-* and no custom base URL is set — native OpenAI is used.
Per-agent base URLs — STRATEGIST_API_BASE_URL, OPERATOR_API_BASE_URL, and EVALUATOR_API_BASE_URL override the global base URL for that agent when set (see table below).

Models whose names do not match built-in providers (Gemini, Claude, Grok, or gpt-*) require a base URL (OPENAI_API_BASE, API_BASE_URL, or the relevant per-agent URL); otherwise initialization fails with a clear configuration error.

Per-Agent Model Overrides

You can override the shared global model for each agent independently.

Agent	Model	API key	Base URL
Strategist	`STRATEGIST_MODEL`	`STRATEGIST_MODEL_API_KEY`	`STRATEGIST_API_BASE_URL`
Operator	`OPERATOR_MODEL`	`OPERATOR_MODEL_API_KEY`	`OPERATOR_API_BASE_URL`
Evaluator	`EVALUATOR_MODEL`	`EVALUATOR_MODEL_API_KEY`	`EVALUATOR_API_BASE_URL`

For Gemini-oriented deployments, a common pattern is to use a stronger model for Strategist and Evaluator work while using a faster model for Operator execution. The exact model names should match the providers available in your environment.

Configuration Profiles

Configuration Pattern	Typical Settings	Best For
Balanced default	`ACCURACY=standard`, `GRANULARITY=adaptive`	General research workflows where you want a good balance of speed, structure, and completeness.
Faster exploration	`ACCURACY=eco`, optionally `GRANULARITY=low`	Early-stage investigation, quick experiments, and situations where you want shorter iteration cycles.
Higher rigor	`ACCURACY=pro`, often with `GRANULARITY=high`	More demanding analyses where explicit planning, stronger verification, and fuller scientific coverage matter more than speed.
Adaptive control (experimental)	`ACCURACY=adaptive` and/or `GRANULARITY=adaptive`	Lets the Strategist and Operator adjust theory level and task breakdown based on workflow complexity instead of fixed presets.

A practical research workflow is to begin with ACCURACY=eco to establish a fast baseline, then switch to higher accuracy mode to refine the result through auto-improvement or a human-guided follow-up pass based on that initial run.

Machine learning in workflows

The Operator is instructed that the Python stack can include scikit-learn (declared in package metadata) and PyTorch (pulled in with the MACE ML potential stack and GPU-oriented images). You can ask QUASAR to fit models, preprocess data, or run small training jobs as part of a research workflow when your environment provides the necessary dependencies.