Yield (start_index, batch) tuples for successive `size`-sized slices.

Create fallback PhaseValidation when validation fails. v1.4.0: Now distinguishes between infrastructure errors (INCONCLUSIVE) and ot

Uses an LLM to filter a list of criteria to only those relevant to the asset type.

Extract decision and rejection reason from context.

Calculate shift duration in hours.

Check if date is a weekend.

Load model photos from preset folder.

Load a JSON or JSONL file. Returns the parsed Python value. For .jsonl, returns a list of one record per non-empty line.

Generate n fake case contexts. case_target is always the target.

Render a JSON array as markdown. Array of objects → each item is a `## {heading}` section separated by `---` so the ch

Convert image bytes to numpy array. Args: image_bytes: Image as bytes Returns: Numpy array (BGR format)

Check if proposed rule conflicts with existing rules. Returns list of warning messages.

Phase 1: Evaluate all rules and collect matches. Enforces scope-based mutual exclusion (at most one match per scope). Do

Derive a deterministic file identifier from a GCS URI. Uses MD5 hex digest so the same file always gets the same ID regardless of content cha

Resizes and center-crops an image to match the dimensions of a video frame. Args: video_frame_bytes: Reference video frame as bytes

Create a frame with a product (colored rectangle) on a background. Args: width, height: Frame dimensions product_rect: (x, y, w,

Helper to create a simple presentation for testing.

Describe a single garment image for VTO purposes. Extracts brand, type, color, material, and all visible logos/text/marks with EXTERIOR/INTER

Detect the percentage of a target color in an image using HSV color space. Args: image_bytes: Image data as bytes. target_rgb: Th

Render and return HTML for done page

Ensures the image artifact exists, creating it if necessary from GCS. If an `image_filename` is provided that starts with "gs://", it will be

Evaluate a single condition against data. Args: data: The full case context (provisional output + intermediate artif

Evaluate the wearing quality of a generated VTO image. No reference garment needed — only assesses how naturally the outfit is worn. Args:

Extract a numbered section (e.g., "0", "4.2", "3.8"). Handles formats: - ## 0. Data Source Priorities - ### 4.2

Fetches regional economic metrics directly from St. Louis Fed (FRED) API. Replaces legacy BigQuery labor tables. Support MSAs like Austin, Ra

Scrapes Tax Foundation for the latest state corporate income tax rates. This bypasses legacy BigQuery dependencies and provides real-time dat

Return (to_extract, dup_stubs). to_extract: files that need a Cloud Run call. dup_stubs: files whose content_hash already maps to a previousl

Format a rule dict for terminal display.

Generate the next message from the agent based on the input message. Args: message: The input message from the user or tool.

Loads tools from the MCP server. @return The list of tools.

Lazily instantiates the ADK App and Agent only when needed.

Creates a Dataproc BatchControllerClient for a specific region. Args: region: The GCP region for the client. Returns: A

Get a configured BigQuery client.

Initializes and returns a GCS client with robust project detection.

Finds the next date that doesn't have a finalized roster. Returns: Date string in YYYY-MM-DD format for the next unscheduled date.

Count non-white pixels in an image as a quality metric. More non-white pixels = more complete shoe (less background). This helps select the m

Get a specific rule group by ID.

Get a specific validation step from a phase's output. Args: phase: Phase key (e.g., "phase1", "phase4")

Get the workspace path using configuration.

Check if a sequence of nodes represents a valid path through the allowed graph. Allows tolerance for temporary misclassifications and oscillation

Match generated frame labels to reference image labels. Args: frame_labels: List of predicted labels for generated frames frame_i

Extract markdown tables from a .xlsx file.

Post-process a generated video by trimming to the most similar end frame. Finds the frame most similar to the target end image and trims the vide

Cheap raw-text extraction without invoking any LLM. Used by the preprocess service to run the relevance classifier on raw text BEFORE the exp

Submit an evaluation task to the shared thread pool. Args: reference_face_bytes: Reference face image as bytes. generated_vto_byt

Perform a set subtraction operation on two binary masks. Returns a mask containing only the regions present in mask_a that are NOT in mask_b.

Generate SQL to update order status in BigQuery using ADK tools. This function now returns SQL for execution by the ADK toolset instead of ex

Upload a single file to Google Cloud Storage.

AuthMiddleware wraps handlers to enforce authentication.

User operations

Load retrieves the data given the reference path.

Append a batch of chunk rows to the staging table.

Apply confidence-based gating to violation determination. v1.4.0: New function to minimize false positives by requiring high confide

Safely applies font style (name, size, color) to a given text run.

Build a human-readable case summary for the SME.

Creates a standard set of headers for GitHub API requests. Includes the 'Accept' and 'User-Agent' headers. If a token is provided, it al

Creates a JSON string conforming to the ContentMap.Contents protobuf structure that the EvalTask expects.

Decodes a base64 string to bytes, returning None if invalid.

Evaluates an iterable of expressions. Args: elts: The AST nodes representing the iterable elements. namespace: The current namesp

Evaluates a list of statements. Args: stmts: The AST nodes representing the statements. namespace: The current namespace.

Find the LibreOffice binary (handles macOS vs Linux paths).

Retrieves a GitHub authentication token from various sources. The function prioritizes sources in the following order: 1. A token passed

Retrieves the user's id to use in the agent's operations, along with whether to continue processing or not.

Check if extracted text has meaningful content.

Runs the LLM as a judge on the given message.

Build a Pub/Sub trigger request body from an expense dict.

Mint an ID token using the default credentials.

Return current time as HH:MM:SS.mmm.

Returns web request result wrapped in BeautifulSoup object Arguments: url (`str`): If no url or html is provided, use the current

Helper to convert a Firestore document to a dictionary.

Run fn(*args) with exponential backoff on transient API errors.

Helper method to run an agent and get the final response.

Run the prompt asynchronously and return the assistant message. Args: new_message: The new message from the user. functio

Add a new rule to the engine (Phase 1: expert-defined). Args: rule_dict: Rule definition dict matching the schema

Check if all required classes are present in the frames. Args: product_position_frames: List of classified positions. Returns:

Check if two amount strings match within tolerance.

Converts a list of BigQuery Row objects to a JSON string, handling datetime objects by converting them to ISO 8601 strings.

Calculates a score based on the evaluation result. Args: evaluation_result: The result of the media evaluation. Returns: An

Check if a time slot is available (no conflicting events).

Return True if more than one person is detected in the response.

Create a new image where the masked area is visible and the rest is white. The image is cropped to the mask's bounding box to center the extracte

Count the number of faces detected in a Vision API response.

Detect the background color by sampling pixels from the borders of the frame. Returns the median color value (BGR format). Args: fra