Create 4K canvases for images, stacking the last two if there are 4 images. For 1-3 images: Creates individual canvases with consistent height sc

Build full prompt with investigator guidelines prepended.

Lazily initializes Selenium driver to avoid import-time startup failures.

Load HRIS data.

Parse JSON/dict input from LLM, handling all common formats.

Save HRIS data.

Generates a compact text-based progress bar.

Detect velcro/hook-and-loop fasteners in shoe images.

Evaluate all conditions (AND-joined). Returns: (all_passed, evaluation_details)

Evaluate how faithfully the glasses in the generated image match the reference. Args: client: Gemini client instance generated_im

Generate text content using Gemini. Args: content_pieces: List of text strings, image/doc bytes, or GCS paths. client: Gemini cli

Create and configure the ADK BigQuery toolset. Uses Application Default Credentials for authentication.

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

Check if LangChain Gmail toolkit is available and initialized.

Check if a sequence of nodes represents a valid path through the allowed graph.

ADK Agent Callback that intercepts user input before the agent runs.

ADK Agent Callback that intercepts model output after the LLM runs.

Parse a formatted amount string like '1,100.00' into a float.

Parses a timestamp string into an integer representing total seconds. Supports formats: raw seconds ('322'), MM:SS ('05:22'), or HH:MM:SS ('

POST with exponential backoff on 5xx, 429, and connection errors. - 5xx + ConnectionError + Timeout: retried up to `attempts` times, backof

Upload bytes directly to Google Cloud Storage.

Evaluates an assignment statement (e.g., x = 1). Args: evaled_value: The value to assign. target: The AST node representing the t

Build a description block for VTO prompts from garment descriptions. Args: garment_descriptions: List of dicts with "general" and "detail

Pick a heading for an item from common identifier fields.

Load JSON file, return empty dict if not found.

Parse a dollar string like '$150,000' into a float.

Render a JSON object as markdown. Scalar fields → `**key**: value` lines. Nested objects/arrays → sub-section under their own heading.

Resolve dynamic field references in condition values. If the value is a string matching the pattern _DYNAMIC_<field_path>_,

Per-run staging name: prevents two concurrent pipeline runs from clobbering each other. If a run crashes the table is left behind under its ru

Calculate SSIM similarity between two images. Args: img1_bytes: First image as bytes img2_bytes: Second image as bytes Retur

Create canvases for multiple images with consistent height scaling. Calculates the maximum safe height that fits all images within the margins,

Create an image of a solid color.

Determine whether a tool execution should be halted (pure function).

Evaluate all reference garments against the generated VTO image in parallel. Uses the product fitting's evaluate_garment (0-10 scale) with EX

Evaluate if the person in a generated image matches the reference face. This is the main function to use. It handles: 1. Cropping face from g

Extract a specific step (e.g., "15", "22", "33"). Handles formats: - ### Step 15: Duplicate Invoice Detection -

Extract product from background and upscale it. Performs background removal, upscaling, and optionally reapplies the scaled original mask to

Generate an image with the person placed in a new background using 2-step approach. Args: client: Gemini client instance preproce

Generates a visual from a single string prompt using a multi-modal model. Returns the GCS URI of the generated image or a temp local path if

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

Return a fresh-enough ID token for the given audience. Thread-safe. Caches token per audience and refreshes 10 min before the 60-min expiry.

Generate the prompt used for interpolation video generation with a generic product title. Args: client: Gemini client for title generatio

Return True if person count is acceptable (0 or 1).

Generate the next ALF-NNN rule ID.

Preprocess images with optional upscaling and canvas creation. Used by both Interpolation and R2V modes for video generation. Args:

Process a list of frame bytes, detect product boundaries, and crop to maximize product size. Returns the processed frames as a bytelist at the sp

Search the product catalog for items matching a text query. Uses vector similarity search to find products matching the description. Args:

Build the SQL query to find files needing chunking. Two modes: - Incremental (default): files re-extracted since last successful index,

Create three framings from a full-body image: lower body, upper body, face. Matches the animation flow: video starts on the lower body, reveals u

Determine risk tier and base rate from underwriting results. Returns: Tuple of (risk_tier_name, base_interest_rate)

Determines if the asset is a video or image based on MIME type.

Return the current METRICS dict.

Load JSON file, return empty dict if not found.

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

Truncate a value for display.

Check if the mask contains exactly one distinct object. This is useful for discarding empty masks (0 contours) or fragmented masks (2+ contou

Classify product type and each image's view angle in a single Gemini call. Returns: dict: { "product_type": "shoes" | "cars"

cleanupData removes data created during tests.

Fetches HUD FMR with Title Case key matching for FY2026. Args: county_fips: 5-digit County FIPS code or common city name (e.g. "

Find the index of the frame most similar to a reference frame. Compares frames against the reference frame using Structural Similarity Index

Find the initial and ending original indices where the starting class appears for the second time. If not second appearance return None None.

Forward-fill null/None labels with the previous non-null label. All frames between two consecutive labeled frames get assigned the label of the f

Generate a background mask using Vertex AI image segmentation. Automatically resizes large images before segmentation to stay within API limits.

Helper to get config from env or user input.

Fixture to provide a mock environment with tools and policy.

Return list of rule dicts from rule_base.json.

Calculate the Structural Similarity Index (SSIM) between two images. Args: frame1_bytes: First image as bytes frame2_bytes: Secon

Reads a JSON/JSONL string and returns a schema representation of all objects.

Fetch the latest order with 'order_placed' status from BigQuery using ADK tools. This is a wrapper function that uses the ADK execute_sql too

Get nested value from dictionary

Return a standard GCS client. Transfer Manager handles connection pooling internally, so custom adapters are rarely needed.

Select images for closure classification based on desirable views. Args: bytes_list: List of image bytes. images_classes: List of

Validate video based on people counts per frame. A video is valid if there is no more than 1 person from a given point onwards. Returns the i

Reads instruction text from a file relative to this script.

Merge multiple MP4 videos (in bytes) into a single video using temporary files. Apply a specific speed to each clip. Args: videos_by

Map action to the corresponding page

Execute the full cleanup step. Main entry point, called by the Vertex AI Pipeline cleanup component. It performs: 1. **Detect** — Query

Generate product fitting images for catalogue enrichment (B2B). Product fitting is designed for enriching product catalogues: it takes a sing

Sorts a list of (mask, value) tuples based on the mask's horizontal position. The score is calculated by summing the transformed x-coordinates of

Stack two images horizontally side by side. Resizes both images to have the same height (using the smaller height) while maintaining aspect r

DSN constructs the PostgreSQL Data Source Name.

Appends a value to a list in the session state.