Components

Input NumPy File

Loads NumPy array files from the bound directory and emits each as a Tensor, either whole or sliced into batches along a chosen axis at a paced rate. Use it to replay saved arrays into a tensor pipeline.

Extract Text Embedding (BGE)

@baai

Encodes each String into a dense Embedding with a bound BGE sentence-transformers checkpoint, with an optional retrieval query prefix and L2 normalisation. Suited to text retrieval, clustering, similarity, and RAG query/passage encoding.

Visualize Zone Transitions

@bleenco

Zone-transition overlay. Draws a configured polygon on Image and labels each [(Rectangle, UInt64)] tuple as Entry (code 1), Exit (code 2), or box-only; emits annotated Image. Pair with detect_zone_transition.

Input Audio File

Decodes the bound audio clip and plays it out as a real-time stream of AudioFrame chunks, looping the playlist when asked. Use it for offline batch processing or deterministic, repeatable audio benchmarks.

Input Text File

Reads text files from the bound directory and emits each one whole as a single String, optionally looping. Accepts any text-shaped corpus, from Markdown and JSON to CSV, SQL, and log files.

Branch

Conditional pass-through gate: takes a Bool predicate and a value of any type t, and emits Maybe — the value wrapped as present when the predicate is true, or absent when false — preserving the stream cadence on both arms.

Translate Text (HuggingFace)

@huggingface

Translates a String or [String] with a HuggingFace seq2seq model loaded in-process (default NLLB-200), forcing the target-language decode token, and returns a string, string list, or language-tagged record chosen from the resolved downstream port.

Output Bytes HTTP

Forwards each Bytes value verbatim over independent egress channels — a binary WebSocket broadcast and/or async HTTP POST as application/octet-stream — so raw binary blobs from an upstream encoder reach live clients or an external endpoint.

Subtract Background (OpenCV)

@opencv

Per-frame Image foreground extractor built on a stateful Gaussian-mixture or K-nearest-neighbour background model. Emits oneof t[Image, Mask] — a one-channel foreground intensity frame, or a boolean foreground mask — with the arm fixed by the wired downstream.

Estimate Face Embeddings (AuraFace v1)

@fal

Face identity embedding extractor: takes an Image and a [[Landmark]] of per-face keypoints, aligns each face by a five-point affine crop, runs an ONNX recognition network, and emits one L2-normalised [Embedding] per face in input order.

Answer Question (HuggingFace)

@huggingface

Extracts the answer to a question from a passage with a HuggingFace QA reader: takes a question String and a context String, finds the span of the context that answers it (default DeBERTa-v3-large SQuAD2.0), and emits the answer String or a QAAnswer record tagged with its char offsets + score.

Stabilize Video (OpenCV)

@opencv

Online video stabiliser that takes shaky Image frames and emits steadied Image frames. It estimates inter-frame motion, smooths the warp over a moving window, and re-warps each frame, optionally cropping the margins the warp reveals to keep output size constant.

Classify Text (HuggingFace)

@huggingface

Runs a fine-tuned text-classification model over a String and emits ranked [Classification], a single Classification, or raw [Double] scores in the model's label order, by output slot, when a trained model exists (sentiment, topic, intent).

Classify Speech Emotion (Dimensional)

@audeering

Regresses continuous valence-arousal-dominance affect from each AudioFrame with a wav2vec2 dimensional model, emitting a {arousal: Double, dominance: Double, valence: Double} record, a raw [Double] triple, or [Analytics.Metric].

Rerank Text (BGE)

@baai

Cross-encoder reranking of retrieved passages. Scores every query-document pair in a RerankInput with a local BGE cross-encoder, sorts by descending relevance, optionally truncates, and emits [RerankResult] carrying each document, score and original index.

Lift Reroll

Re-rolls a UInt64 count plus N element streams (each $ts) back into N parallel lists (each [$ts]) of that length, buffering count element-ticks before emitting one tick. Inverse of lift_unroll, whose length channel supplies the count.

Detect Objects Open Vocabulary (Florence 2)

@microsoft

Localizes whatever a per-message String query phrase names within an Image via Florence-2 open-vocabulary detection, projecting the detections onto the connected output shape: boxes, geometry, classes, a count, or an overlay. Offline from the bound model.

Detect Objects by Text Prompt (Ultralytics YOLO World)

@ultralytics

Embeds a list of free-text class names once at startup with YOLO-World's text encoder, then detects only those categories in each Image and projects the detections onto the connected output shape: boxes, geometry, classes, box+class tuples, a count, or an overlay.