Extraction Tables

Extraction Tables let you build structured grids where AI reads each of your documents and fills in the cells. Define the columns you care about — sample size, methodology, primary outcome, confidence intervals — and Virza extracts that data from every paper in your table automatically.

What extraction tables are for

Extraction tables solve the problem of reading dozens of papers and manually copy-pasting data points into a spreadsheet. Common use cases:

Each cell is extracted by AI from the paper’s full text — not from the abstract alone. The model reads the methods, results, and discussion sections to find the most accurate answer.

Requirements

Before running extraction on a document, it must meet both of these conditions:

1. Fully processed — the document must have a Ready status. Documents that are still scanning or parsing do not have extracted text yet, so the AI has nothing to read.
2. Has readable text — scanned images with no OCR, password-protected PDFs, and corrupt uploads cannot be extracted. Virza runs OCR automatically, but very low-quality scans may produce poor results.

Creating a table

Defining columns

Columns are the backbone of your table. Each column has:

• Name — a short label shown in the table header (e.g., “Sample size”)
• Prompt — the instruction the AI follows when reading each paper (e.g., “What is the total number of participants in the study? Return only the number.”)
• Data type — the expected format of the extracted value

Column data types

Choosing the right data type helps the AI format its answer correctly and makes the table easier to read at a glance.

Writing effective prompts

The quality of your extraction depends almost entirely on your prompts. A vague prompt produces vague results.

Rules for good prompts:

1. Be specific about what you want — instead of “Sample size”, write “What is the total number of participants enrolled in the study? Include dropouts. Return only the integer.”
2. Specify the format — “Return only the number”, “Return yes or no”, “Return a comma-separated list”
3. Tell the AI where to look — “As reported in the Methods section”, “As stated in Table 1 or the Results section”
4. Handle missing data — “If not reported, return N/A”
5. Avoid ambiguity — if a paper could have multiple answers (e.g., multiple arms with different sample sizes), specify which you want: “Total enrolled across all arms combined”

Prompt examples by column type:

Column limits

Each table supports up to 20 columns. If you need more, consider splitting your extraction into multiple focused tables (e.g., one table for study design, another for outcomes).

Using pre-built template columns

Template packs give you pre-written column definitions with proven prompts. To use them:

1. Open your table and click Add Column
2. Switch to the Templates tab
3. Select a template pack and choose the columns you want to add
4. Click Add Selected Columns

You can modify a template column’s name or prompt after adding it. The template key is retained for reference but does not affect extraction.

RCT template columns

Cohort template columns

Qualitative template columns

Adding and removing documents

Adding more documents after creation

Click Add Documents from the table view. Any new documents are added with Pending cells for all existing columns. Run extraction again to fill the new rows.

Removing documents

Click the row’s context menu and select Remove Document. This removes the document from the table but does not delete it from your library. Cells are soft-deleted (retained internally for 30 days) and can be re-added.

Running extraction

Re-extracting cells

Extraction is not destructive. You can re-run extraction at any time:

• Re-extract all — click the Extract button again; Virza re-processes only cells that are still Pending or Failed, leaving already Done cells untouched.
• Re-extract a single cell — click the cell, then click Re-extract this cell from the cell detail panel.
• Re-extract after editing a prompt — if you update a column’s prompt, cells already marked Done are reset to Pending for that column only, then re-extraction fills them with the new prompt.

Understanding “Not Found” results

A Not Found result means the AI read the full document and could not identify the information you asked for. Common causes:

Not Found cells count as a valid result and can be sorted and filtered like any other.

Editing the table

Renaming the table

Click the table title to edit it inline. Press Enter or click away to save.

Editing column prompts

Click the column header, then Edit Column. Update the prompt and save. All cells for that column are reset to Pending and will be re-extracted on the next extraction run.

Reordering columns

Drag column headers left or right to reorder them. Column positions are saved automatically.

Deleting columns

Click the column header → Delete Column. Cells are soft-deleted (recoverable within 30 days). This cannot be undone from the UI within the same session.

Exporting to CSV

Click Export → Download CSV from the table toolbar. The CSV file contains:

• One row per document
• One column per extraction column, plus document metadata columns (title, authors, year, DOI)
• Cell values as plain text (lists are joined with semicolons)
• Empty cells for Not Found or pending values

The filename is extraction-table-{table-id}.csv. You can open it directly in Excel, Google Sheets, or import it into R/Python for statistical analysis.

Table limits

Tips and best practices

Start narrow, then expand — begin with 5–10 high-priority columns on a small subset of papers (10–15). Review the quality before running on your full corpus.

Use collections to organise documents — create a collection for the papers relevant to your review before opening Extraction Tables. You can filter the document picker by collection to add them all at once.

Test prompts on a single paper first — before running on 80 documents, test your column prompt on one paper you know well. Verify the cell captures what you expect, then scale up.

Number columns are strict — if a paper reports a range (“100–150 participants”), the AI may return the range as text. Make your prompt explicit: “If a range is given, return the midpoint as an integer.”

Not all documents process equally — preprints, conference papers, and theses may use different section structures. If you get many Not Found cells for a specific paper, check its document status and consider whether it was fully parsed.

Re-extract is non-destructive — if you tweak a prompt and run extraction again, only the affected column’s cells are reset. All other cells remain intact.

Troubleshooting

The table is stuck in “Extracting” status

The extraction worker may have crashed or the job may have timed out. Refresh the page and click Re-extract. If it persists, check that your documents are all in Ready status.

All cells show “Not Found” for a specific document

The document likely failed to parse correctly. Go to your library, find the document, and check its status. If it shows a processing error, try re-uploading the original file.

Cells show “Failed” for many documents

This usually indicates an issue with the AI service (transient error). Click Re-extract to retry. Failed cells are re-queued automatically on re-extraction.

The Extract button is disabled or missing

Check that you have Editor role or higher in the workspace. Viewers can read and export tables but cannot trigger extraction or add columns. See Roles & Permissions for details.

The Extraction Tables option is not visible in the sidebar

This feature requires a Pro plan and the research_extraction_tables feature flag to be enabled for your workspace. Contact your workspace owner or check Billing & Plans.

Privacy and data handling

• Document text is sent to Virza’s AI gateway (Virza Cortex) for extraction. No document content is used to train models.
• Extraction jobs are workspace-scoped — cells from one workspace are never visible to another.
• Extracted cell values are stored encrypted in Virza’s database.
• Deleting a table soft-deletes all cells. Hard deletion occurs automatically after 30 days.

For full privacy details, see Data Isolation.