Required prompt structure

MetaPaperLens needs your prompt to instruct the AI model to emit two specific JSON structures: an evidence array and an extraction_confidence object. The viewer reads these to render PDF highlights and per-block confidence badges. Without them the extraction will still run, but those panels stay empty.

If your prompt doesn't yet ask for these structures, copy the examples below into your prompt body. Substitute the placeholder field names (<your_field>, <data_block_1>) with the actual top-level keys your prompt asks the model to produce.

1. The evidence array

Every sample / result object must carry an evidence array. Each entry has exactly four keys: snippet, page, source, and field. The field value is a JSON path identifying which extracted value the evidence supports — formatted to mirror the JSON shape your prompt asks for.

"evidence": [
  {
    "snippet": "TABLE 1. Descriptive statistics by group...",
    "page": 3,
    "source": "Table 1",
    "field": "samples[0].<your_field>"
  },
  {
    "snippet": "the sample comprised 147 non-clinical adolescents",
    "page": 2,
    "source": null,
    "field": "samples[0].sample_id"
  }
]

Rules:

• snippet must be verbatim from the paper — character-for-character, no paraphrase, no ellipses.
• page is the sequential PDF page number (1 = first page), as an integer. Never omit.
• source is the table / figure identifier (e.g. "Table 2") or null.
• field is the JSON path into your output, e.g. "samples[0].demographics._table[0]".

2. The extraction_confidence object

A top-level extraction_confidence object (sibling of samples, summaries, etc.) with one entry per major extracted-data block in your output. Each entry has a level rated high / medium / low, plus a notes string when not high.

"extraction_confidence": {
  "<data_block_1>": { "level": "high" },
  "<data_block_2>": {
    "level": "medium",
    "notes": "country inferred from author affiliations; not stated explicitly"
  },
  "<data_block_3>": {
    "level": "low",
    "notes": "values reported only for the overall scale; subgroup entries estimated"
  }
}

Rules:

• level ∈ {"high", "medium", "low"}.
• notes (≤ 200 chars) is required on medium and low; optional on high.
• Exactly once at the top level — not inside samples, not inside evidence, not per-record.
• One entry per major data block — not for evidence itself or for extraction_confidence itself.

Fix-my-prompt checklist

1. Find the JSON schema or output-shape section in your prompt.
2. Add an evidence array alongside your main data field, with the four-key spec above.
3. Add an extraction_confidence object at the top level of the output, with one entry per major data block.
4. Paste the worked examples from this page into your prompt so the model sees the structure it must emit.
5. Return to the prompt editor — the warning should clear once both blocks are present.

Why this matters: the renderer matches each evidence entry's field path back to the JSON you produced, then highlights the corresponding page region. The confidence object drives the per-block badges on the results panel. Both depend on the structures being present and JSON-shaped — keyword mentions alone aren't enough.