Abstractive Compression Prompt with Factuality Constraints

This prompt is for AI engineers and operators who need to compress long evidence passages into shorter, fluent summaries when extractive snippets are too verbose or disjointed for the target context window. The core job-to-be-done is reducing token cost and latency without introducing claims that are absent from the source material. The ideal user is someone integrating this into a production RAG pipeline, an agent workflow, or a document processing system where downstream tasks depend on factual accuracy. You should have the original source text, a clear compression target (e.g., token count or reduction ratio), and a defined downstream consumer—whether that's another model, a human reviewer, or a structured data extraction step.

Do not use this prompt when the evidence is short enough to pass through verbatim, when extractive snippets with citations are required for compliance, or when the compression ratio is so aggressive that essential facts will inevitably be dropped. It is also the wrong tool when the downstream task requires exact legal or medical terminology that must not be paraphrased. In regulated domains, always pair this prompt with a factuality verification step and a human review gate before the compressed output is used in decisions. The prompt is designed for abstractive compression, meaning it will rephrase and condense—it will not produce a word-for-word extract. If you need verbatim quotes mapped to claims, use an extractive snippet selection prompt instead.

Before wiring this into production, define your factuality eval criteria: compare the compressed output against the source for additions, omissions, and semantic drift. Common failure modes include hallucinated dates, invented proper nouns, and subtle shifts in causality or certainty. Start with a moderate compression ratio (e.g., 50%) and tighten only after validating that your eval pass rate holds. If the compression introduces unsupported claims, adjust the constraints, add few-shot examples of faithful compression, or reduce the target ratio. The next section provides the copy-ready template with placeholders for your source, constraints, and output schema.

This template is the core instruction set for an abstractive compression task. It is designed to be copied directly into your prompt management system, IDE, or orchestration code. The prompt forces the model to produce a fluent, condensed summary while anchoring every factual claim to the source text. It is not a general summarization prompt—it is specifically hardened for factuality by requiring the model to separate certain facts from uncertain interpretations and to explicitly mark any necessary inferences.

textCopy
You are a precise evidence compressor. Your task is to produce a condensed, fluent summary of the provided source text. The summary must be significantly shorter than the original while preserving all claims critical to the user's task.

# SOURCE TEXT
[SOURCE_TEXT]

# USER TASK CONTEXT (OPTIONAL)
[TASK_CONTEXT]

# COMPRESSION TARGET
Target length: approximately [TARGET_LENGTH] [UNIT: words | sentences | tokens].

# OUTPUT FORMAT
Return a valid JSON object with the following schema:
{
  "summary": "The condensed abstractive summary string.",
  "certain_facts": ["List of factual claims directly and unambiguously supported by the source."],
  "uncertain_inferences": ["List of any necessary interpretations, implications, or connections you made that are not explicitly stated in the source. Include 'NONE' if empty."],
  "omitted_topics": ["List of major topics or entities present in the source that were intentionally omitted from the summary due to length constraints."]
}

# ABSOLUTE CONSTRAINTS (VIOLATING THESE IS A FAILURE)
1. DO NOT introduce any fact, figure, name, date, or event that is not present in the source text.
2. DO NOT express an opinion on the source material.
3. If the source is too short or empty to compress, set 'summary' to an empty string and explain why in 'uncertain_inferences'.
4. Every item in 'certain_facts' must be a verbatim or near-verbatim extraction from the source.
5. If you are unsure whether a claim is explicit, place it in 'uncertain_inferences', not 'certain_facts'.

To adapt this template, replace the bracketed placeholders with your runtime values. [SOURCE_TEXT] is the only strictly required input; without it, the prompt should be aborted before inference. [TASK_CONTEXT] is optional but highly recommended—it guides the model on which information is salient for the downstream task, preventing it from compressing away critical evidence. [TARGET_LENGTH] and [UNIT] let you enforce a specific budget. The JSON output schema is non-negotiable for production use because it separates the summary from the factuality audit trail (certain_facts and uncertain_inferences), which is essential for automated hallucination checks. Do not remove the 'ABSOLUTE CONSTRAINTS' section; these are the behavioral contract that prevents the model from treating this as a creative summarization task.

Before integrating this prompt into a production pipeline, test it against a golden dataset of source texts where you know the expected certain facts. A common failure mode is the model placing a true but slightly paraphrased fact into uncertain_inferences due to an overly strict interpretation of 'verbatim'. You may need to adjust the constraint wording for your specific model. After generation, always run a programmatic check: diff the certain_facts list against the source text to ensure every item is a substring or a near-exact match. Any fact that fails this check should be treated as a potential hallucination and either moved to uncertain_inferences or flagged for human review.

The abstractive compression prompt is not a standalone tool; it is a component inside a larger context engineering pipeline. In production, you will typically call this prompt after retrieval and before the final generation step. The harness must accept the original source text and any task-specific constraints, inject them into the template placeholders, and then validate the compressed output before it reaches downstream consumers. Because this prompt is designed to reduce token costs while preserving factual fidelity, the harness must treat the compression step as a lossy but auditable transformation. Always store the original source text alongside the compressed output so that factuality checks and citation verification can be performed later.

Wire the prompt into your application with a pre-inference validation layer that checks: (1) the source text is non-empty and within the model's maximum context window, (2) the compression ratio target is a sensible number between 0.1 and 0.9, and (3) any domain-specific terminology lists are properly formatted as JSON arrays or comma-separated strings. After the model returns the compressed summary, run a post-inference factuality check using a separate verification prompt (such as the Factuality Preservation Check Prompt After Compression from this content group) that compares the compressed output against the original source. If the factuality check flags additions, omissions, or semantic drift above a configurable threshold, trigger a retry with a stronger constraint instruction or escalate for human review. For high-stakes domains like healthcare or legal, require human approval before the compressed context is used in any downstream generation step.

Choose your model based on the compression ratio and latency requirements. For aggressive compression (ratios below 0.3), use a model with strong instruction-following and long-context handling, such as Claude 3.5 Sonnet or GPT-4o. For moderate compression in high-throughput pipelines, a smaller model like Claude 3 Haiku or GPT-4o-mini may suffice, but always validate factuality more aggressively with smaller models. Implement exponential backoff retries for transient API failures, but cap retries at three attempts for factuality failures to avoid infinite loops. Log every compression call with: the original source length in tokens, the compressed output length, the compression ratio achieved, the factuality check score, and whether human review was triggered. This telemetry will help you tune the compression aggressiveness over time and detect model drift. Finally, never pass the compressed output directly to end users without an explicit review step if the source material contains regulated content, conflicting evidence, or time-sensitive claims that could expire between compression and use.