Extraction features: complete breakdown

What extraction stores

Extraction rows store structured values for included or eligible studies. They can represent baseline variables, outcome measurements, and study-design or study-characteristics fields.

• Study and group/arm identifiers.
• Outcome or variable names.
• Numerical values such as mean, SD, events, totals, sample size, confidence intervals, p-values, and related fields where the template supports them.
• Units, timepoints, subgroup paths, notes, and custom field values.
• Source page, quote, source-document reference, and provenance metadata.

PDF handling

• Direct PDF upload is available.
• Open-access PDF fetch can be used where the project has the required feature access and the source is available.
• PDF access follows project membership and role permissions.
• The evidence viewer keeps the report beside the extraction workspace.

Evidence viewer

The evidence viewer helps reviewers inspect the source report while extracting. Use it to locate the table, paragraph, or figure where a value came from.

• Available text, page, region, and table information can support evidence capture.
• Selected text can be sent to extraction fields where the workflow supports it.
• Project-wide PDF search is premium-gated and only shown when enabled.
• Source quotes and pages should be captured for values that affect analysis or manuscript tables.

MCP extraction suggestions

When a project is connected to Coventra MCP, an AI agent can help with extraction from uploaded full-text PDFs. Coventra keeps the task tied to the protocol: the agent works against the outcomes and fields already configured in the project rather than freely inventing what to extract.

The agent receives controlled evidence excerpts from the PDF and proposes values with source anchors that reviewers can inspect in Coventra. This supports the needle-in-a-haystack part of extraction without turning the AI into the final data owner.

MCP suggestions appear as reviewable, read-only proposal rows in the extraction sheet. They do not become canonical extraction data until a researcher accepts them. A separate MCP extraction benchmark article reports the latest workflow results without exposing implementation details.

• No configured outcomes means discovery mode only: the agent may suggest outcomes, but should not create final extraction rows.
• Configured outcomes guide extraction by study, outcome, and relevant comparison or arm.
• Time-to-event rows require an effect estimate plus uncertainty; binary rows require both events and denominator.
• Every submitted value must point back to a source location in the current PDF.
• Repeat submissions are protected so the review queue does not fill with duplicate suggestions.

Extraction grid

• Edit baseline, outcome, and study-design rows.
• Add, update, or remove rows where supported.
• Track source/provenance at the field level.
• Use validation to catch missing or inconsistent values.
• Use version history and restore controls when changes need audit review.

Harmonization helpers

Coventra includes helpers for common evidence-synthesis conversions, but users must verify the result.

• Unit harmonization.
• IQR, range, SE, CI, and p-value conversions where the input shape supports them.
• Median-to-mean conversion in analysis workflows.
• Outcome synthesis and cluster naming support where configured.

Blind dual extraction

Blind extraction support allows independent extraction drafts, conflict comparison, resolution, and reviewer completion summaries. Use it when the protocol requires independent duplicate extraction.

AI suggestion limits

Premium projects can use optional study-characteristics suggestions. Suggestions are not final data. The reviewer must inspect, accept, edit, or reject them.

Data freeze

Data freeze creates a final-analysis snapshot boundary. Use it when extraction and validation are complete enough for final analysis and manuscript exports.

• Freeze after resolving major extraction conflicts.
• Freeze after validation warnings are reviewed.
• Keep a note of what changed if data are edited after a freeze.