coreSR.AI
Learn

Data Collection

Source-first, PDF-aware data extraction with bidirectional provenance

Upload study PDFs and let GLASS — CoreSR's benchmarked, locally hosted extraction model — turn the report into structured, citation-backed study data. Reviewers keep control: accept, edit, or reject every value, preserve source context, and export only the records that are ready for synthesis.

How It Works

Six stages from PDF upload to analysis-ready export, with extracted values tied to source context, model provenance, and reviewer decisions.

1UploadPDFsSingle or bulkUp to 100 files2AIExtractGLASSStructured fields3ReviewFieldsQuotes + contextConfidence signals4FigureWorkflowVision detectorKM Cor / calibration5Verify +CorrectInline editingAudit trail6Export+ AnalyzeXLSX / CSVPairwise / NMA

Benchmarked for Production Extraction

GLASS, CoreSR's locally hosted extraction model, is benchmarked on small expert-derived held-out sets before production use — internal checks, not published external validation.

Held-out accuracy
94%
engine-scored accuracy on an internal held-out extraction set
Value recall
0.92
production harness on a 29-record held-out gold set
Parse failures
0
zero parse failures across the held-out harness run
Model cost
$0
GLASS runs on CoreSR infrastructure — no per-call API cost

No digit hallucination was observed in the production harness run: numeric values are copied verbatim from the source and paired with a supporting quote, so reviewers can verify each value against the paper before it enters an analysis dataset. These are internal benchmarks on small expert-derived sets, not published external validation.

Source-First vs. Question-First

Traditional extraction forms ask questions; this module turns source reports into reviewer-controlled structured records. The difference matters.

Question-first (traditional)

A form asks you a series of questions: "What was the treatment group N?", "What was the HR?", "What was the p-value?" You search the paper for each answer, copy values into a spreadsheet, and hope every context decision is remembered later.

  • ✗No visual link between extracted data and its source
  • ✗Easy to mis-type a number and never notice
  • ✗Hard to audit or re-verify later

Source-first (Data Collection)

The AI reads the report against your schema first, then presents structured values with source context and confidence signals. Reviewers work from proposed evidence into an analysis-ready dataset instead of starting from a blank grid.

  • Every value keeps source context
  • Corrections are logged with full context
  • Re-verification is one click away

Bidirectional Provenance

Every extracted value keeps source context, model provenance, reviewer state, and edit history. Visual crops are retained where figure or table workflows expose page coordinates.

PDF PAGE 4Table 2 — Outcomes0.72 [0.58-0.89]review value← or click bboxEXTRACTION GRIDOutcomeNHR [95% CI]0.72 [0.58-0.89]Every value stored withpage + bbox + textprovenance record

Review value source

Source quotes and PDF context stay next to extracted values so reviewers can verify the interpretation before accepting it.

Audit visual evidence

Figure and chart workflows preserve page-image crops and calibration metadata for source-level review.

Stored with provenance

Extraction records preserve model, timestamp, source context, reviewer action, and corrected value history.

What Gets Extracted

The AI scan classifies every detected source into one of four categories.

Tables

Baseline demographics, outcomes, adverse events, subgroups, sensitivity analyses, dosing, biomarkers. Extracted as structured fields with evidence context and reviewer state.

DemographicsOutcomesAdverse eventsSubgroups

In-text results

Quantitative results stated in running text: hazard ratios, odds ratios, medians, percentages, p-values. Extracted into structured fields (estimate, CI lower, CI upper, p-value, group labels) with the source sentence preserved verbatim.

HR 0.72 [0.58–0.89]Median PFS 14.2 mop=0.002

Figures

Figures containing numeric data (Kaplan-Meier with at-risk tables, forest plots, bar charts with values). KM Cor auto-calibrates Kaplan-Meier crops from printed ticks; other figure types require reviewer calibration.

KM with at-riskForest plotsBar charts

Supplement material

Uploaded separately as supplement documents. Extracted independently, then reviewed alongside the main-article values with their own provenance.

eTable 1eTable 2Subgroup analyses

Workflow Step-by-Step

From upload to analysis-ready export.

STEP 1

Create a Study Project

Each Data Collection project represents one study. Give it a label (e.g. "Chen et al. 2024"), optionally add PMID or DOI. For reviews with many studies, use Bulk Upload to create one project per PDF in a single drop.

Tip:Bulk Upload accepts up to 100 PDFs at once. Filename becomes the default study label; rename anytime.
STEP 2

Upload the PDF

Drop in the main article PDF. The system extracts the full text via pdf-parse (for in-text search and the text-layer highlight fallback), counts pages, and stages the file on disk. Supplements go in separately as additional documents.

Tip:Scanned image-only PDFs work too, but lose the text-layer search/highlight features.
STEP 3

Run AI Extraction

Trigger the single-stage extraction pass. GLASS, CoreSR's locally hosted model, reads the study report against the project schema and returns structured values, source quotes, value types, confidence signals, and provenance.

Tip:Define fields clearly before running extraction. Units, timepoints, eligible effect measures, and expected categories improve downstream review speed.
STEP 4

Review Extracted Fields

The study workspace separates proposed values, evidence quotes, reviewer edits, confidence, and acceptance state. Reviewers decide which values are ready for synthesis.

Tip:Prioritize low-confidence values, unusual effect measures, and values drawn from figures or supplements.
STEP 5

Inspect Tables and Figures

Structured text extraction handles table and in-text values. Figure detection and digitization are available as a separate vision workflow for Kaplan-Meier curves, forest plots, bar charts, line plots, box plots, scatter plots, and flow diagrams.

Tip:KM Cor fits and validates printed axis ticks automatically; reviewer anchors can override that fit. Other chart types still require saved reviewer calibration.
STEP 6

Verify and Correct

Edit any extracted value before accepting it. The original AI value, reviewer correction, source context, timestamp, and user action remain visible in the audit trail.

Tip:Focus verification effort on low-confidence values, statistic-type labels, unusual timepoints, and values pulled from figures or supplements.
STEP 7

Map Columns (optional)

For records that will feed into pairwise / NMA analyses, confirm each field maps to a standard role such as outcome, treatment_group, events_t, n_t, effect_estimate, ci_lower, ci_upper, or p_value.

Tip:You only need analysis mappings for records you plan to synthesize; descriptive study characteristics can remain as reviewed extraction fields.
STEP 8

Export

Export reviewed extraction records to formatted XLSX, analysis-ready CSV, and downstream module handoffs with provenance and reviewer state preserved.

Tip:The provenance log is your PRISMA-compliant audit trail. Save it with your extraction records.

Benchmarked Extraction, Verification, and Audit Trail

Extraction is fast, but it remains reviewer governed. Three safeguards keep the workflow accountable.

Single-stage extraction

The active text extraction path is a single pass through GLASS, CoreSR's locally hosted model. A single high-performing stage keeps output behavior simpler, faster, and easier to audit.

Reviewer verification

Reviewers accept, edit, reject, or leave extracted values unresolved. Final analysis datasets are built from reviewer-confirmed records, not raw model output.

Append-only audit log

Every edit preserves the original AI value, reviewer change, timestamp, source context, and model provenance. The system keeps enough history to re-check extraction decisions later.

At a Glance

The capabilities and limits that shape how you use the module.

Max file size50 MB per PDF
Bulk uploadUp to 100 PDFs at once
Primary extraction modelGLASS — CoreSR's locally hosted model (single-stage)
Internal benchmark accuracy94% engine-scored accuracy on a held-out extraction set (internal benchmark)
Production harness check0.92 value recall on a 29-record held-out gold set; zero parse failures, no digit hallucination observed
Benchmark cost$0 API cost — GLASS runs on CoreSR infrastructure
Export formatsXLSX (formatted), CSV (analysis-ready), provenance log
Audit trailReviewer-attributed edits, acceptance state, source quotes, and model provenance
Figure extractionKM Cor auto-calibration for KM curves; reviewer-calibrated workflow for other charts

Export and Integration

Extracted data flows into the rest of the CoreSR workflow without re-typing.

Formatted XLSX

Structured extraction workbook with reviewer-confirmed values, source context, corrections, and provenance metadata for audit.

Analysis-ready CSV

Confirmed outcome records flatten into standard fields: study, outcome, treatment_group, events_t, n_t, effect_estimate, ci_lower, ci_upper, p_value, and effect_measure.

Direct to analysis

Data Collection projects can be linked to a Pairwise or NMA project at creation time. The analysis module reads the confirmed extraction records as a dataset with no intermediate file step.

Video Tutorial Coming Soon

A walkthrough video demonstrating bulk upload, AI extraction, reviewer verification, figure handling, and export is in production.

Frequently Asked Questions

Common questions about PDF-aware data extraction.

How is this different from manual data extraction?
Traditional extraction asks you to hunt through each paper field by field. Data Collection starts with the uploaded PDF and produces structured, citation-backed extraction proposals for study characteristics, arms, outcomes, effect estimates, counts, percentages, and continuous outcomes. Reviewers then accept, edit, or reject each value with the source context preserved.
How is extraction performed?
Extraction runs as a single pass through GLASS, CoreSR's locally hosted model, using a project-stable schema, Zod validation, type coercion, confidence handling, and full per-call provenance. Figure and chart extraction are handled separately by the vision pipeline because visual digitization has different failure modes from text extraction.
How accurate is AI extraction?
On internal benchmarks, GLASS scores 94% engine-scored accuracy on a held-out data extraction set, and the production extraction harness holds 0.92 value recall on a 29-record held-out gold set with zero parse failures and no digit hallucination observed — numeric values are copied verbatim from the source with a supporting quote. These are internal benchmarks on small expert-derived sets, not published external validation. This is still reviewer-governed extraction: every proposed value remains editable and auditable before downstream analysis.
What if the AI misses a value?
You can add or edit values during review. The AI pass is meant to accelerate first-pass extraction, not lock the dataset; reviewer-confirmed corrections become the authoritative record for export.
Can I extract data from supplementary materials?
Yes. Upload the supplement as a separate document in the same project. The extractor processes supplement documents independently and keeps supplement-derived values separate for reviewer verification.
What happens when I correct a value?
The original AI-extracted value is preserved alongside the reviewer correction, source context, timestamp, and edit state. Nothing is silently overwritten; the audit trail keeps enough context to explain how the final dataset was produced.
Can I export the data directly into pairwise / NMA?
Yes. Confirmed outcome records can be exported with standard analysis fields such as outcome, treatment group, events, n, effect estimate, confidence interval, p-value, and effect measure. Those files can feed Pairwise and NMA workflows directly.
Can I upload many PDFs at once?
Yes. The "Bulk Upload" button on the dashboard accepts up to 100 PDFs simultaneously — each file becomes its own study project, with the filename as the default label. You can then work through them one by one, triggering scans and extractions as you go. This is the recommended workflow for a full systematic review.
How do bidirectional highlights work?
Extracted values retain source context, model provenance, reviewer state, and edit history. Where the figure/table workflow has page-image coordinates, the system also preserves the page and crop metadata needed for visual audit.
How long does extraction take?
Extraction typically averages well under one minute per study. Real project timing depends on PDF length, supplement volume, and queue load, and because GLASS runs on CoreSR infrastructure there is no per-call API cost.
Is my data private?
PDFs are stored on the CoreSR server, not shared publicly. Primary text extraction runs on GLASS on CoreSR-controlled infrastructure. Vision-based figure digitization is a separate external-provider path; KM Cor sends the selected figure crop to Gemini Pro. Extracted data, corrections, and audit records remain in the CoreSR database. You can delete a project at any time; documents, extractions, and audit records cascade on delete.
What PDF formats are supported?
Native text-layer PDFs from journals and preprint servers work best. Scanned PDFs and figure-heavy supplements can still be reviewed through the vision pipeline, but reviewer verification becomes more important because OCR quality, rotated pages, and low-resolution charts can affect extraction quality. File size is capped at 50 MB per PDF.

Ready to start?

Create a study, bulk-upload your PDFs, and let GLASS prepare citation-backed records so your team can focus on verification.

References

1. Higgins JPT, Thomas J, Chandler J, et al. Cochrane Handbook for Systematic Reviews of Interventions, Chapter 5: Collecting data. 2nd ed. Wiley; 2019.

2. Li T, Higgins JPT, Deeks JJ (eds). Chapter 5: Collecting data. In: Cochrane Handbook for Systematic Reviews of Interventions version 6.4. Cochrane; 2023.

3. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. doi:10.1136/bmj.n71

4. Büchter RB, Weise A, Pieper D. Development, testing and use of data extraction forms in systematic reviews: a review of methodological guidance. BMC Med Res Methodol. 2020;20:259.

CoreSR — AI-powered evidence synthesis