What Is Cohesion?

What Is Cohesion?

Cohesion in natural-language processing is the linguistic property of how the parts of a text are tied together. It comes from explicit signals: pronoun reference (“the model… it… its”), conjunctions (“however”, “therefore”, “as a result”), lexical repetition or synonyms across sentences, and discourse markers that link clauses. For LLM output, cohesion is the surface-level glue that makes a paragraph readable; coherence is the deeper logical consistency. FutureAGI surfaces cohesion as one component inside text-quality evaluators alongside coherence, conciseness, and helpfulness — a model can produce cohesive nonsense, so it is rarely scored alone.

Why Cohesion matters in production LLM systems

A response can be factually correct and still hard to read. Low cohesion shows up as choppy paragraphs, missing pronoun antecedents, abrupt topic shifts, and dropped discourse markers. Users react by re-reading, asking the same question twice, or escalating to a human — even when the underlying answer was right. The pain is invisible in eval suites that only score factual accuracy or relevance.

Roles feel this differently. A product manager sees thumbs-down feedback labeled “confusing” and “hard to follow” with no clear evaluator owning the issue. A developer notices that prompt changes meant to shorten output broke pronoun reference — the new version drops “it” and “this” without antecedents. A compliance reviewer reading regulated outputs (legal, medical, financial) finds passages where the text reads well but has no logical thread, indicating the model retrieved or generated facts without organizing them.

In agent and multi-turn settings, cohesion compounds. An agent that produces five tool-call summaries with no transition between them looks like a list of disconnected facts. A summarization step that drops back-references to the user’s original question reads as off-topic. In 2026 RAG and agent stacks, cohesion is part of the user-experience layer that separates demos from products: factually sound, well-glued responses retain users; factually sound but choppy responses produce thumbs-down at the same rate as wrong ones.

How FutureAGI handles Cohesion

FutureAGI does not isolate cohesion as a single product metric — that would be too narrow. The closest direct surface is the CoherenceEval framework evaluator inside fi.evals, which scores logical and surface flow together. For dedicated text-quality work, teams pair CoherenceEval with SummaryQuality, IsConcise, and AnswerRelevancy to triangulate. A response that scores high on Faithfulness but low on CoherenceEval is factually correct but poorly glued — the right action is prompt tuning, not retrieval changes.

A real workflow: a documentation team running a RAG agent on traceAI-langchain notices thumbs-down rate at 11% despite Groundedness averaging 0.94. They sample 200 thumbs-down responses into a dataset and run CoherenceEval plus SummaryQuality. Cohesion-related failures cluster at long responses (>400 tokens) where the model loses pronoun reference. They use PromptWizard to optimize the system prompt with explicit instructions about back-reference and re-run regression eval — CoherenceEval rises from 0.71 to 0.88 with no drop in factual scores.

For custom rubrics — for example “every paragraph must reference the previous paragraph’s subject” — CustomEvaluation wraps a judge-model prompt as a callable evaluator with score, label, and reason. FutureAGI’s approach is rubric-first, not single-metric, because cohesion is one dimension of a multi-dimensional quality target.

How to measure or detect Cohesion

Useful signals to surface cohesion problems:

• CoherenceEval: framework-level evaluator that returns a coherence-and-flow score; a strong proxy for cohesion in LLM output.
• SummaryQuality: catches summaries where bullets are correct but unconnected.
• IsConcise: low scores often correlate with rambling outputs that also lose cohesion.
• AnswerRelevancy: cohesive outputs that drift off-topic still fail relevance.
• Thumbs-down rate by response length: cohesion failures cluster at long outputs; bucket by token count.
• User-feedback proxy: comments tagged “confusing” or “hard to follow” without “wrong” indicate a cohesion or coherence gap.

Minimal Python:

from fi.evals import CoherenceEval, AnswerRelevancy

coh = CoherenceEval().evaluate(
    input="Explain how RAG handles long documents.",
    output=model_response,
)
rel = AnswerRelevancy().evaluate(
    input="Explain how RAG handles long documents.",
    output=model_response,
)
print(coh.score, rel.score)

Common mistakes

• Conflating cohesion with coherence. A paragraph can be tightly glued and logically wrong, or logically tight and choppy; treat them as separate axes.
• Relying on token-overlap metrics. BLEU and ROUGE do not measure pronoun reference or discourse links; they reward surface n-gram match.
• Optimizing only for cohesion. Rewarding well-glued text without checking factual accuracy produces fluent hallucinations.
• Scoring only short outputs. Cohesion problems concentrate in long-form generation; sample by length.
• Skipping a baseline. Compare candidate prompt versions against a held-out cohort, not against ad-hoc examples.