Scientific Paper Quote Extraction Prompt

This prompt is built for research synthesis tool builders who need to extract exact, verbatim quotes from scientific papers to support claims, populate evidence databases, or generate citation-aware summaries. The ideal user is an engineering lead or AI product developer integrating this prompt into a pipeline where factual fidelity and omission rate are the primary quality metrics. The required context includes a full paper text (often noisy from PDF extraction), a specific claim or question to ground, and an output schema that enforces section provenance. Use this prompt when you need precise evidence spans with their original section context, not paraphrased summaries or general topic extraction.

The prompt handles the unique challenges of academic text: dense domain-specific terminology, section-based document structure (Abstract, Methods, Results, Discussion), PDF extraction artifacts like broken words and merged columns, and references to figures and tables that cannot be visually inspected. It is designed to operate in pipelines where downstream users will verify extracted quotes against original sources, making verbatim accuracy non-negotiable. The prompt is not suitable for generating literature reviews, summarizing paper contributions in plain language, or extracting claims without their supporting evidence spans. It also should not be used when the input is a paper abstract alone—full text is required for section-level provenance.

Before deploying this prompt, ensure your pipeline includes PDF text extraction that preserves section boundaries, a validation step that checks quote fidelity against the source, and a human review queue for low-confidence extractions. The prompt works best with structured output constraints that enforce quote text, section name, and paragraph location for every extracted span. Avoid using this prompt for papers behind paywalls where full text is unavailable, or for domains where the cost of a missed quote (omission) is higher than the cost of an irrelevant extraction. For those scenarios, pair this prompt with a recall-oriented retrieval step and a separate precision filter.

The prompt below is designed to extract claims, findings, and methodology details from scientific papers as verbatim quotes with section context. It handles common PDF extraction artifacts, figure references, and multi-section document structures. Copy this template and replace each square-bracket placeholder with your specific inputs before wiring it into your research synthesis pipeline.

textCopy
You are an evidence extraction assistant for scientific papers. Your task is to extract verbatim quotes from the provided paper text that directly support or describe specific claims, findings, or methodology details.

## INPUT
Paper text:
[PAPER_TEXT]

Extraction targets:
[EXTRACTION_TARGETS]

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "extractions": [
    {
      "quote": "verbatim text from the paper",
      "section": "section name where the quote appears",
      "target_match": "which extraction target this quote supports",
      "relevance": "direct_support | partial_support | contextual",
      "artifacts_detected": ["list of PDF artifacts if any, e.g., hyphenation_break, garbled_equation, missing_figure_reference"],
      "confidence": 0.0_to_1.0
    }
  ],
  "missing_targets": ["targets with no supporting quote found"],
  "artifact_notes": "summary of significant extraction artifacts encountered"
}

## CONSTRAINTS
- Extract quotes verbatim, preserving original wording, spelling, and punctuation.
- Do not paraphrase, summarize, or alter the quoted text.
- Flag any PDF extraction artifacts that may affect quote accuracy (hyphenation breaks, garbled special characters, missing Greek letters, corrupted equation fragments, figure reference placeholders).
- For each extraction target, provide the single most relevant quote. If multiple quotes are relevant, select the most specific one.
- If a figure or table is referenced but not available as text, note this in artifacts_detected and reduce confidence accordingly.
- If no quote supports a target, list it in missing_targets with a brief explanation.
- Section names should reflect the paper's actual section headers where possible.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]

Adaptation guidance: Replace [PAPER_TEXT] with the full or sectioned text of the scientific paper, including any preprocessing to normalize line breaks and remove header/footer noise. [EXTRACTION_TARGETS] should contain a structured list of the specific claims, findings, or methodology details you need quotes for—be explicit about what constitutes a match. [FEW_SHOT_EXAMPLES] should include 2-3 examples of correct extractions from similar papers, showing how to handle artifacts and section attribution. Set [RISK_LEVEL] to high if the extracted quotes will be used in published research synthesis or regulatory submissions, which triggers additional validation requirements in the implementation harness.

What to do next: After adapting the template, wire it into your extraction pipeline with validation checks for quote fidelity, section accuracy, and artifact flagging. Run the prompt against a golden dataset of papers with known ground-truth quotes to measure factual fidelity and omission rate before production use. For high-risk applications, always include a human review step to verify extracted quotes against the original PDF, especially when artifacts are detected.

The Scientific Paper Quote Extraction Prompt is designed to be called from an application layer that has already retrieved or ingested a scientific paper and split it into manageable chunks. The prompt expects a chunk of paper text as [PAPER_CHUNK] and a specific extraction target defined in [EXTRACTION_TARGET] (e.g., 'claims', 'methodology details', 'findings'). The application is responsible for chunking the paper by section boundaries where possible, preserving section headers as context, and passing each chunk through the prompt. Do not send an entire paper in a single call; chunk size should stay within the model's context window while leaving room for the prompt instructions and the expected output.

The implementation harness must include a validation layer that checks the structured output against the defined [OUTPUT_SCHEMA] before the extracted quotes are stored or displayed. Each extracted quote object must contain the required fields: quote_text (verbatim string), section_context (the paper section it came from), extraction_type (claim, finding, methodology, etc.), and confidence (a score or label). The validator should reject any quote where quote_text cannot be found verbatim in the source chunk—this is the primary defense against hallucinated quotes. If the model returns a quote that fails the verbatim substring check, the harness should log the failure, increment a retry counter, and re-invoke the prompt with the same chunk and an added [CONSTRAINTS] note emphasizing exact text fidelity. After three failed retries, flag the chunk for human review rather than silently accepting a paraphrased or fabricated quote.

Model choice matters for this workflow. Models with strong instruction-following and long-context fidelity (such as Claude 3.5 Sonnet or GPT-4o) perform better at preserving exact quote boundaries. Avoid smaller or older models that tend to paraphrase even when instructed to extract verbatim text. If you are processing papers with significant PDF extraction artifacts—such as broken words, hyphenation remnants, or merged columns—preprocess the text to normalize whitespace and dehyphenate before passing it to the prompt. The prompt includes handling instructions for figure references and artifact-affected text, but cleaner input produces more reliable extractions.

Logging and observability are critical for scientific use cases where extraction fidelity is auditable. Log every prompt invocation with: the chunk ID, paper DOI or identifier, extraction target, the raw model output, validation results, and any retry attempts. Store these logs in a structured format that allows downstream analysis of extraction quality by paper, section, or extraction type. If the application surfaces extracted quotes to users, always display the section_context alongside the quote so readers can navigate to the source. For high-stakes research synthesis, consider a human review queue for quotes flagged with low confidence or those that failed initial validation before being included in any published output.

When integrating this prompt into a larger RAG or research pipeline, the extraction step should run after document ingestion and chunking but before any summarization or synthesis step. The extracted quotes become the grounded evidence set that downstream answer-generation prompts can cite. This ordering ensures that synthesis is built on verified, verbatim source material rather than on model-recalled approximations of what the paper says. Avoid the temptation to skip extraction and ask a single prompt to both read and summarize—separating extraction from synthesis is the architectural decision that makes grounding auditable.