summarizer

name	summarizer
version	1.0.0
description	Summarises a block of text to a target word count using extractive techniques.
inputs	[{"name":"text","type":"string","description":"The text to summarise (plain text, up to ~10 000 words)","required":true},{"name":"max_words","type":"integer","description":"Target maximum word count for the summary (default 100)","required":false},{"name":"language","type":"string","description":"BCP-47 language code for the output summary, e.g. 'en', 'es' (default: 'en')","required":false}]
outputs	[{"name":"summary","type":"string","description":"The condensed summary text"},{"name":"word_count","type":"integer","description":"Actual word count of the returned summary"},{"name":"compression_ratio","type":"number","description":"Ratio of summary words to input words (lower = more compressed)"}]
capabilities	["stateless","streaming"]
dependencies	["httpx>=0.27.0"]
resources	{"vcpu":1,"memory_mib":128,"boot_timeout_ms":500}

Overview

The summarizer agent condenses large blocks of text into concise summaries. It uses a sentence-scoring heuristic (TF-IDF-style term frequency) to select the most representative sentences from the input, staying within the requested max_words limit.

This agent is deliberately stateless — each request is independent.

Algorithm

1. Tokenise the input into sentences (split on ., !, ? followed by whitespace).
2. Score each sentence by the frequency of its constituent words across the full document.
3. Select top-scoring sentences in document order until max_words is reached.
4. Return the selected sentences joined by spaces.

Example

Request:

{
  "text": "The unikernel-fabric is a spec-driven runtime...",
  "max_words": 50
}

Response:

{
  "summary": "The unikernel-fabric is a spec-driven runtime for deploying autonomous agents.",
  "word_count": 12,
  "compression_ratio": 0.24
}

Error Handling

• If text is empty or fewer than 10 words, return HTTP 422 with {"error": "text too short to summarise"}.
• If max_words is less than 5, clamp to 5.
• If the detected or requested language is not supported, fall back to "en" and note it in a warning field.