Configuration

The frozen config dataclasses (GazeConfig, SearchConfig, ImageProcessingConfig, and related) plus config_context(), a ContextVar-based helper for task-scoped overrides without mutating globals. For the available settings and usage patterns, see Configuration.

config

Configuration classes for GAZE.

Centralized, frozen configuration for the constants, limits, and tunable parameters used across the framework.

ImageProcessingConfig dataclass

Configuration for image processing operations.

Attributes:

Name	Type	Description
min_image_size	int	Minimum dimension for images and crops (pixels)
max_image_dimension	int	Maximum allowed width or height in pixels
min_zoom_factor	float	Minimum allowed zoom factor
max_zoom_factor	float	Maximum allowed zoom factor
min_contrast_factor	float	Minimum allowed contrast factor
max_contrast_factor	float	Maximum allowed contrast factor
min_threshold_window	int	Minimum intensity window width for threshold tool. Prevents destructive narrow windowing that destroys diagnostic info.
min_window_width	int	Minimum intensity window width for the window_level tool (distinct from the threshold tool's min_threshold_window).
default_jpeg_quality	int	Default JPEG quality for encoding (1-100)
max_tool_encode_dimension	int	Longest side, in pixels, of an image sent back to the model in a tool result. The ImageManager keeps the full-resolution image for subsequent operations and measurements; this only bounds the base64 payload, which otherwise reaches megabytes after successive zooms and is re-sent every turn.

Source code in src/gaze/config.py

@dataclass(frozen=True)
class ImageProcessingConfig:
    """Configuration for image processing operations.

    Attributes:
        min_image_size: Minimum dimension for images and crops (pixels)
        max_image_dimension: Maximum allowed width or height in pixels
        min_zoom_factor: Minimum allowed zoom factor
        max_zoom_factor: Maximum allowed zoom factor
        min_contrast_factor: Minimum allowed contrast factor
        max_contrast_factor: Maximum allowed contrast factor
        min_threshold_window: Minimum intensity window width for threshold tool.
            Prevents destructive narrow windowing that destroys diagnostic info.
        min_window_width: Minimum intensity window width for the window_level
            tool (distinct from the threshold tool's min_threshold_window).
        default_jpeg_quality: Default JPEG quality for encoding (1-100)
        max_tool_encode_dimension: Longest side, in pixels, of an image sent
            back to the model in a tool result. The ImageManager keeps the
            full-resolution image for subsequent operations and measurements;
            this only bounds the base64 payload, which otherwise reaches
            megabytes after successive zooms and is re-sent every turn.
    """

    min_image_size: int = 10
    max_image_dimension: int = 16384
    min_zoom_factor: float = 0.5
    max_zoom_factor: float = 4.0
    min_contrast_factor: float = 0.5
    max_contrast_factor: float = 3.0
    min_threshold_window: int = 50
    default_jpeg_quality: int = 85
    max_tool_encode_dimension: int = 1024
    min_brightness_factor: float = 0.5
    max_brightness_factor: float = 2.0
    min_sharpness_factor: float = 0.1
    max_sharpness_factor: float = 3.0
    max_grid_divisions: int = 8
    min_gaussian_sigma: float = 0.5
    max_gaussian_sigma: float = 5.0
    max_morphological_iterations: int = 5
    min_clahe_clip_limit: float = 1.0
    max_clahe_clip_limit: float = 4.0
    max_clahe_tile_size: int = 32
    min_window_width: int = 50

    def __post_init__(self) -> None:
        if self.min_image_size < 1:
            raise ValueError(f"min_image_size must be >= 1, got {self.min_image_size}")
        if self.max_image_dimension < self.min_image_size:
            raise ValueError(
                f"max_image_dimension ({self.max_image_dimension}) "
                f"must be >= min_image_size ({self.min_image_size})"
            )
        if self.min_zoom_factor >= self.max_zoom_factor:
            raise ValueError(
                f"min_zoom_factor ({self.min_zoom_factor}) "
                f"must be < max_zoom_factor ({self.max_zoom_factor})"
            )
        if self.min_contrast_factor >= self.max_contrast_factor:
            raise ValueError(
                f"min_contrast_factor ({self.min_contrast_factor}) "
                f"must be < max_contrast_factor ({self.max_contrast_factor})"
            )
        if not 1 <= self.min_threshold_window <= 255:
            raise ValueError(
                f"min_threshold_window must be between 1 and 255, got {self.min_threshold_window}"
            )
        if self.max_tool_encode_dimension < self.min_image_size:
            raise ValueError(
                f"max_tool_encode_dimension ({self.max_tool_encode_dimension}) must be "
                f">= min_image_size ({self.min_image_size})"
            )
        if not 1 <= self.default_jpeg_quality <= 100:
            raise ValueError(
                f"default_jpeg_quality must be between 1 and 100, got {self.default_jpeg_quality}"
            )
        if self.min_brightness_factor >= self.max_brightness_factor:
            raise ValueError(
                f"min_brightness_factor ({self.min_brightness_factor}) "
                f"must be < max_brightness_factor ({self.max_brightness_factor})"
            )
        if self.min_sharpness_factor < 0:
            raise ValueError(f"min_sharpness_factor must be >= 0, got {self.min_sharpness_factor}")
        if self.min_sharpness_factor >= self.max_sharpness_factor:
            raise ValueError(
                f"min_sharpness_factor ({self.min_sharpness_factor}) "
                f"must be < max_sharpness_factor ({self.max_sharpness_factor})"
            )
        if not 2 <= self.max_grid_divisions <= 20:
            raise ValueError(
                f"max_grid_divisions must be between 2 and 20, got {self.max_grid_divisions}"
            )
        if self.min_gaussian_sigma <= 0:
            raise ValueError(f"min_gaussian_sigma must be > 0, got {self.min_gaussian_sigma}")
        if self.min_gaussian_sigma >= self.max_gaussian_sigma:
            raise ValueError(
                f"min_gaussian_sigma ({self.min_gaussian_sigma}) "
                f"must be < max_gaussian_sigma ({self.max_gaussian_sigma})"
            )
        if not 1 <= self.max_morphological_iterations <= 20:
            raise ValueError(
                f"max_morphological_iterations must be between 1 and 20, "
                f"got {self.max_morphological_iterations}"
            )
        if self.min_clahe_clip_limit >= self.max_clahe_clip_limit:
            raise ValueError(
                f"min_clahe_clip_limit ({self.min_clahe_clip_limit}) "
                f"must be < max_clahe_clip_limit ({self.max_clahe_clip_limit})"
            )
        if not 2 <= self.max_clahe_tile_size <= 64:
            raise ValueError(
                f"max_clahe_tile_size must be between 2 and 64, got {self.max_clahe_tile_size}"
            )
        if self.min_window_width < 2:
            raise ValueError(f"min_window_width must be >= 2, got {self.min_window_width}")

CacheConfig dataclass

Configuration for caching behavior.

Attributes:

Name	Type	Description
max_cache_size	int	Maximum number of cached entries
cache_duration_seconds	int	Time-to-live for cache entries in seconds
evict_ratio	float	Fraction of cache to evict when over limit (0.0-1.0)

Source code in src/gaze/config.py

@dataclass(frozen=True)
class CacheConfig:
    """Configuration for caching behavior.

    Attributes:
        max_cache_size: Maximum number of cached entries
        cache_duration_seconds: Time-to-live for cache entries in seconds
        evict_ratio: Fraction of cache to evict when over limit (0.0-1.0)
    """

    max_cache_size: int = 500
    cache_duration_seconds: int = 300  # 5 minutes
    evict_ratio: float = 0.5

    def __post_init__(self) -> None:
        if self.max_cache_size < 1:
            raise ValueError(f"max_cache_size must be >= 1, got {self.max_cache_size}")
        if self.cache_duration_seconds <= 0:
            raise ValueError(
                f"cache_duration_seconds must be > 0, got {self.cache_duration_seconds}"
            )
        if not 0.0 < self.evict_ratio < 1.0:
            raise ValueError(
                f"evict_ratio must be between 0.0 and 1.0 exclusive, got {self.evict_ratio}"
            )

SearchConfig dataclass

Configuration for search operations.

Attributes:

Name	Type	Description
timeout_seconds	int	Request timeout in seconds
max_retries	int	Maximum number of retry attempts
rate_limit_delay_seconds	float	Delay between API calls in seconds
max_content_preview_length	int	Maximum characters for content preview
max_snippet_length	int	Maximum characters for snippet extraction
max_content_for_llm	int	Maximum characters for LLM formatting
ncbi_base_url	str	Base URL for NCBI E-utilities API
openi_base_url	str	Base URL for OpenI API

Source code in src/gaze/config.py

@dataclass(frozen=True)
class SearchConfig:
    """Configuration for search operations.

    Attributes:
        timeout_seconds: Request timeout in seconds
        max_retries: Maximum number of retry attempts
        rate_limit_delay_seconds: Delay between API calls in seconds
        max_content_preview_length: Maximum characters for content preview
        max_snippet_length: Maximum characters for snippet extraction
        max_content_for_llm: Maximum characters for LLM formatting
        ncbi_base_url: Base URL for NCBI E-utilities API
        openi_base_url: Base URL for OpenI API
    """

    timeout_seconds: int = 30
    max_retries: int = 3
    rate_limit_delay_seconds: float = 1.0
    max_content_preview_length: int = 500
    max_snippet_length: int = 100
    max_content_for_llm: int = 5000
    ncbi_base_url: str = "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/"
    openi_base_url: str = "https://openi.nlm.nih.gov/api/search"

    def __post_init__(self) -> None:
        if self.timeout_seconds <= 0:
            raise ValueError(f"timeout_seconds must be > 0, got {self.timeout_seconds}")
        if self.max_retries < 1:
            raise ValueError(f"max_retries must be >= 1, got {self.max_retries}")
        if self.rate_limit_delay_seconds < 0:
            raise ValueError(
                f"rate_limit_delay_seconds must be >= 0, got {self.rate_limit_delay_seconds}"
            )
        for attr in ("ncbi_base_url", "openi_base_url"):
            _validate_base_url(
                getattr(self, attr),
                attr,
                allowed_hostnames=_ALLOWED_SEARCH_HOSTNAMES,
            )

AgenticConfig dataclass

Configuration for the multi-turn agentic loop.

Holds the tunable parameters of AgenticProcessorBase. Processor constructor arguments (max_turns, max_tokens, temperature) override the corresponding defaults here for a single processor; values that are not exposed as constructor arguments (the hard turn limit, nudge and idle-tool budgets, tool-content cap) are taken from this config.

Attributes:

Name	Type	Description
max_turns_limit	int	Hard upper bound on max_turns (requests above this are clamped).
default_max_turns	int	Default number of turns when none is given.
default_max_tokens	int	Default completion-token budget per turn.
default_temperature	float	Default sampling temperature (0.0 = greedy).
max_consecutive_nudges	int	Recovery nudges before sending the force-finalize message.
idle_tool_turns_limit	int	Turns with zero tool calls (in agentic mode) before force-finalizing to avoid token waste.
max_tool_content_chars	int	Maximum characters retained from a single tool result before truncation (limits prompt-injection surface).

Source code in src/gaze/config.py

@dataclass(frozen=True)
class AgenticConfig:
    """Configuration for the multi-turn agentic loop.

    Holds the tunable parameters of ``AgenticProcessorBase``. Processor
    constructor arguments (``max_turns``, ``max_tokens``, ``temperature``)
    override the corresponding defaults here for a single processor; values
    that are not exposed as constructor arguments (the hard turn limit, nudge
    and idle-tool budgets, tool-content cap) are taken from this config.

    Attributes:
        max_turns_limit: Hard upper bound on ``max_turns`` (requests above this
            are clamped).
        default_max_turns: Default number of turns when none is given.
        default_max_tokens: Default completion-token budget per turn.
        default_temperature: Default sampling temperature (0.0 = greedy).
        max_consecutive_nudges: Recovery nudges before sending the
            force-finalize message.
        idle_tool_turns_limit: Turns with zero tool calls (in agentic mode)
            before force-finalizing to avoid token waste.
        max_tool_content_chars: Maximum characters retained from a single tool
            result before truncation (limits prompt-injection surface).
    """

    max_turns_limit: int = 30
    default_max_turns: int = 10
    default_max_tokens: int = 16384
    default_temperature: float = 0.0
    max_consecutive_nudges: int = 2
    idle_tool_turns_limit: int = 3
    max_tool_content_chars: int = 8_000

    def __post_init__(self) -> None:
        if self.max_turns_limit < 1:
            raise ValueError(f"max_turns_limit must be >= 1, got {self.max_turns_limit}")
        if not 1 <= self.default_max_turns <= self.max_turns_limit:
            raise ValueError(
                f"default_max_turns ({self.default_max_turns}) must be between 1 and "
                f"max_turns_limit ({self.max_turns_limit})"
            )
        if self.default_max_tokens < 1:
            raise ValueError(f"default_max_tokens must be >= 1, got {self.default_max_tokens}")
        if not 0.0 <= self.default_temperature <= 2.0:
            raise ValueError(
                f"default_temperature must be between 0.0 and 2.0, got {self.default_temperature}"
            )
        if self.max_consecutive_nudges < 1:
            raise ValueError(
                f"max_consecutive_nudges must be >= 1, got {self.max_consecutive_nudges}"
            )
        if self.idle_tool_turns_limit < 1:
            raise ValueError(
                f"idle_tool_turns_limit must be >= 1, got {self.idle_tool_turns_limit}"
            )
        if self.max_tool_content_chars < 1:
            raise ValueError(
                f"max_tool_content_chars must be >= 1, got {self.max_tool_content_chars}"
            )

GazeConfig dataclass

Root configuration for GAZE.

Provides access to all sub-configurations. Can be customized by passing individual config objects or by modifying the defaults.

Example

Use defaults

config = GazeConfig()

Customize specific settings

config = GazeConfig( cache=CacheConfig(max_cache_size=1000), search=SearchConfig(timeout_seconds=60), )

Access settings

print(config.image.max_zoom_factor) # 4.0 print(config.cache.cache_duration_seconds) # 300

Source code in src/gaze/config.py

@dataclass(frozen=True)
class GazeConfig:
    """Root configuration for GAZE.

    Provides access to all sub-configurations. Can be customized by
    passing individual config objects or by modifying the defaults.

    Example:
        # Use defaults
        config = GazeConfig()

        # Customize specific settings
        config = GazeConfig(
            cache=CacheConfig(max_cache_size=1000),
            search=SearchConfig(timeout_seconds=60),
        )

        # Access settings
        print(config.image.max_zoom_factor)  # 4.0
        print(config.cache.cache_duration_seconds)  # 300
    """

    image: ImageProcessingConfig = field(default_factory=ImageProcessingConfig)
    cache: CacheConfig = field(default_factory=CacheConfig)
    search: SearchConfig = field(default_factory=SearchConfig)
    agentic: AgenticConfig = field(default_factory=AgenticConfig)

get_config

get_config() -> GazeConfig

Get the current default configuration.

Returns the context-local override when one is active, otherwise the process-wide default configuration.

Returns:

Type	Description
GazeConfig	The global default GazeConfig instance

Source code in src/gaze/config.py

@beartype
def get_config() -> GazeConfig:
    """Get the current default configuration.

    Returns the context-local override when one is active, otherwise the
    process-wide default configuration.

    Returns:
        The global default GazeConfig instance
    """
    return _ConfigHolder.get()

set_config

set_config(config: GazeConfig) -> None

Set the global default configuration.

Thread-safe replacement of the global configuration.

Parameters:

Name	Type	Description	Default
config	GazeConfig	New configuration to use as default	required

Example

from gaze.config import set_config, GazeConfig, CacheConfig

custom_config = GazeConfig( cache=CacheConfig(max_cache_size=1000) ) set_config(custom_config)

Source code in src/gaze/config.py

@beartype
def set_config(config: GazeConfig) -> None:
    """Set the global default configuration.

    Thread-safe replacement of the global configuration.

    Args:
        config: New configuration to use as default

    Example:
        from gaze.config import set_config, GazeConfig, CacheConfig

        custom_config = GazeConfig(
            cache=CacheConfig(max_cache_size=1000)
        )
        set_config(custom_config)
    """
    _ConfigHolder.set(config)

reset_config

reset_config() -> None

Reset the global configuration to defaults.

Convenience wrapper that restores a fresh GazeConfig(). Useful in test teardown to prevent config leakage between tests.

Source code in src/gaze/config.py

@beartype
def reset_config() -> None:
    """Reset the global configuration to defaults.

    Convenience wrapper that restores a fresh ``GazeConfig()``.
    Useful in test teardown to prevent config leakage between tests.
    """
    _ConfigHolder.set(GazeConfig())

config_context

config_context(config: GazeConfig) -> Iterator[GazeConfig]

Temporarily override configuration for the current thread/task only.

This does not mutate the process-wide default set by :func:set_config. Nested contexts restore correctly, and concurrent threads/tasks keep their own overrides.

Parameters:

Name	Type	Description	Default
config	GazeConfig	Temporary configuration to use inside the block.	required

Yields:

Type	Description
GazeConfig	The temporary configuration (same object as config).

Example

from gaze.config import config_context, GazeConfig, CacheConfig

with config_context(GazeConfig(cache=CacheConfig(max_cache_size=1000))): # get_config() returns the temporary config here ...

original config is restored here

Source code in src/gaze/config.py

@contextlib.contextmanager
@beartype
def config_context(config: GazeConfig) -> Iterator[GazeConfig]:
    """Temporarily override configuration for the current thread/task only.

    This does not mutate the process-wide default set by :func:`set_config`.
    Nested contexts restore correctly, and concurrent threads/tasks keep their
    own overrides.

    Args:
        config: Temporary configuration to use inside the block.

    Yields:
        The temporary configuration (same object as *config*).

    Example:
        from gaze.config import config_context, GazeConfig, CacheConfig

        with config_context(GazeConfig(cache=CacheConfig(max_cache_size=1000))):
            # get_config() returns the temporary config here
            ...
        # original config is restored here
    """
    token = _ConfigHolder.set_context(config)
    try:
        yield config
    finally:
        _ConfigHolder.reset_context(token)