Summarize Text (HuggingFace)

How it fits

Typical backends

Caveats

• I/O contractInput is a single String document and output is a single String summary; there is exactly one summary per input, never a drop. Empty or whitespace-only input returns an empty summary without running the model.
• AccuracySummarization is abstractive: the model paraphrases and may introduce wording not present verbatim in the source, so the summary is not a literal extract and can occasionally state facts the source does not support.
• Hard constraintInput longer than the model's encoder window (model `max_position_embeddings`, e.g. 1024 tokens for facebook/bart-large-cnn) is truncated to its leading window before summarizing; the tail is not seen. Chunk long documents upstream when every section matters, or load a long-context summarizer (e.g. allenai/led-base-16384) via model.
• Parameter interactionmax_length, min_length, and num_beams are re-read every tick and apply on the next summary with no reload; min_length is clamped to never exceed max_length.
• Latencymodel is hot-swappable at runtime and triggers a model + tokenizer reload (and a GPU memory reclaim on CUDA) on the next call, producing a one-call latency spike. device is read once at startup and requires a redeploy to change.
• FallbackA fault inside summarization is logged and degraded to an empty summary; every input emits exactly one output. device starting with `cuda` silently falls back to CPU when CUDA is unavailable, which is impractical for a large summarizer at stream cadence.