Footnote and Annotation Extraction from Tables Prompt

This prompt is designed for data engineers, compliance analysts, and audit teams who need to extract footnotes and annotations from tables embedded in financial filings, regulatory reports, and legal documents. Footnotes in these documents often qualify cell values, disclose assumptions, indicate restatements, or reference external sources. Missing a footnote connection can lead to misinterpretation of the data and downstream errors in reports, dashboards, and automated decisions. Use this prompt when you have already extracted the table body and need to capture the footnote layer: the marker symbols, the full footnote text, and the precise mapping back to the cells they annotate. This prompt assumes the table content and surrounding footnote text are available as structured or semi-structured text input. It does not perform OCR or initial table extraction.

The ideal input is a pre-extracted table structure—such as a JSON array of rows and cells—paired with the full text block of footnotes located below or beside the table. The prompt works best when footnote markers are consistent (e.g., superscript numbers, letters, or symbols like *, †, ‡) and the footnote text follows a predictable pattern of marker followed by explanation. Do not use this prompt when the table extraction itself is unreliable, when footnotes are embedded as images without OCR, or when the document lacks explicit marker-to-text linkage. In those cases, you need upstream table extraction, OCR, or document structure parsing before this prompt can add value. This prompt is also not a substitute for human review in high-stakes filings where a missed footnote could have material regulatory or financial consequences.

Before deploying this prompt, verify that your pipeline can isolate the table region and the footnote region as separate text blocks. If footnotes are interleaved with body text or split across page breaks, you'll need a pre-processing step to reassemble them. After running the prompt, validate the output by checking that every footnote marker in the table has a corresponding entry in the extracted annotations, and that no footnote text is mapped to a non-existent marker. For regulated use cases, route low-confidence mappings to a human reviewer and log the prompt's input, output, and validation results for auditability. The next section provides the copy-ready prompt template you can adapt for your document processing pipeline.

This prompt template is designed to extract footnote markers and their corresponding annotation text from tables embedded in financial, regulatory, or technical documents. It goes beyond simple text extraction by mapping each marker (e.g., superscript numbers, asterisks, daggers) to the specific table cell it annotates, then linking that marker to the full footnote text found elsewhere on the page or in a dedicated notes section. The output is a structured, auditable record that prevents the common failure mode of losing qualifying information when tables are ingested into databases or analytics pipelines.

textCopy
You are an expert document analyst specializing in financial and regulatory table extraction. Your task is to extract all footnotes and annotations from the provided table, map each marker to its source cell, and link it to the full annotation text.

## INPUT
Table image or structured text: [TABLE_DATA]
Surrounding page text or footnote section: [PAGE_CONTEXT]

## INSTRUCTIONS
1. Identify every footnote marker within the table cells (e.g., superscript numbers, letters, asterisks, daggers, or other symbols).
2. For each marker, record the exact cell location using row and column headers or a grid coordinate (e.g., "Row 3, Column '2023 Revenue'").
3. Locate the corresponding full annotation text in [PAGE_CONTEXT]. Match markers precisely; do not guess.
4. If a marker's annotation cannot be found in [PAGE_CONTEXT], flag it as "unresolved" with a confidence score of 0.
5. Classify each annotation by type: [ANNOTATION_TYPES], e.g., "qualification", "assumption", "restatement", "definition", "source", "other".
6. Do not paraphrase or summarize the annotation text. Extract it verbatim.

## OUTPUT_SCHEMA
Return a single JSON object with the following structure:
{
  "table_identifier": "string, a unique label for the table (e.g., 'Table 3: Revenue by Segment')",
  "extraction_confidence": "number between 0 and 1, overall confidence in the extraction",
  "annotations": [
    {
      "marker": "string, the exact marker symbol or text (e.g., '1', '*', '†')",
      "cell_location": {
        "row_index": "number, 0-based row index",
        "column_index": "number, 0-based column index",
        "row_header": "string or null, the row label if available",
        "column_header": "string or null, the column label if available"
      },
      "annotation_text": "string, the verbatim footnote text",
      "annotation_type": "string, one of the types listed in [ANNOTATION_TYPES]",
      "is_resolved": "boolean, true if the annotation was found in [PAGE_CONTEXT]",
      "confidence": "number between 0 and 1, confidence in this specific annotation mapping"
    }
  ],
  "unresolved_markers": [
    {
      "marker": "string",
      "cell_location": { "row_index": "number", "column_index": "number", "row_header": "string or null", "column_header": "string or null" },
      "reason": "string, brief explanation of why it could not be resolved"
    }
  ]
}

## CONSTRAINTS
- Do not invent footnote text. If the annotation is not in [PAGE_CONTEXT], mark it as unresolved.
- Preserve the exact wording, punctuation, and formatting of the annotation text.
- If a single marker references multiple annotations (e.g., '1,2'), create a separate entry for each.
- If the table has no footnotes, return an empty annotations array and set extraction_confidence to 1.

To adapt this prompt for your pipeline, replace the square-bracket placeholders with actual data before each inference call. [TABLE_DATA] should contain the table in a structured format (Markdown, HTML, or a cropped image passed to a multimodal model). [PAGE_CONTEXT] must include the full text of the page or the dedicated 'Notes to Financial Statements' section where the footnote text resides; incomplete context is the primary cause of unresolved markers. [ANNOTATION_TYPES] should be customized to your domain—a financial analyst might use 'accounting policy', 'restatement', and 'estimate uncertainty', while a clinical reviewer might use 'exclusion criterion', 'adverse event', and 'protocol deviation'. After extraction, always run a validation step that checks for orphaned markers and compares the count of extracted annotations against the count of markers visible in the source table.

This prompt is designed as a single step in a larger table extraction pipeline. It assumes you have already isolated a table region from a PDF page and extracted its cell-level content with positional metadata. The prompt takes that structured table data plus the surrounding page text as input and returns a linked annotation structure. Do not send raw PDF bytes or OCR text dumps directly to this prompt—it requires pre-parsed table cells with row/column coordinates and marker positions to function correctly.

Wire the prompt into your pipeline after table cell extraction and before data ingestion. The input should include the full table grid with cell text and coordinates, plus a text buffer of 200-300 words above and below the table to capture footnotes that may appear outside the table boundary. Validate the output against three checks before accepting it: (1) every footnote marker found in table cells must have a corresponding entry in the annotations array, (2) no annotation should reference a marker that does not appear in the table, and (3) each annotation must include a non-empty source_text field that can be traced back to the input context. If any check fails, retry once with the validation errors appended to the prompt as feedback. If the retry also fails, route the table to a human review queue with the original table image crop and the partial extraction results.

For model selection, use a model with strong instruction-following and structured output capabilities. GPT-4o and Claude 3.5 Sonnet both handle the multi-step reasoning required to match superscript markers, roman numerals, asterisks, and other footnote symbols to their definitions. Set temperature to 0 or near-zero to minimize marker mismatches. If your pipeline processes high-volume financial filings, consider caching the prompt prefix that contains the output schema and instructions, as it will remain static across thousands of table extractions. Log every extraction with the input table ID, output annotation count, validation pass/fail status, and any human review decisions for auditability. For regulated documents, retain the raw model response and validation report alongside the final structured output.