Vann

Vann is a literature and evidence workspace, not a search-result summary page. Begin with a focused review question, define the scope and eligibility criteria, and write down the sources or study types you intend to include. A small explicit protocol helps reviewers explain why a paper entered or left the evidence set.

Create an investigation, name the question, and import a source or use discovery. Europe PMC and Crossref can support online discovery when the workspace has internet access. Results may include abstracts or source access limitations; inspect those labels and do not assume every result includes the full text.

Add PDFs, DOCX, XLSX, text, or supported bibliographic records to the source library. Before accepting an import, check its identity, citation, license, extraction status, and page references. A digital PDF can expose searchable text directly; a scan may require optical character recognition. Vann bundles English OCR for a single scanned page or image at a time.

After extraction, compare passages against the original source. Charts, figures, chemical structures, and scanned material still need visual review. OCR can be helpful for locating a passage, but an OCR transcript should not be treated as proof that the extraction is perfect.

In the evidence studio, select the exact passage that supports a claim or measurement. Keep the passage, page or section, source identity, and interpretation together. Candidate claims are rule-based suggestions and need human verification. Vann separates the statement, measurement, interpretation, evidence strength, stance, and review reason so those judgments do not collapse into one label.

When you capture a quantitative measurement, preserve its unit, sample context, condition, comparator, and source location. Import tools validate common measurement tables, but you should still check whether values are comparable and whether a source reports a derived or directly measured quantity.

Use the research map to inspect connections among sources, claims, authors, topics, and entities. A network shows relationships in the records; it does not establish causation or the quality of a body of evidence. Contradiction detection produces candidates for review, not definitive scientific conflicts.

Pooling tools can combine compatible measurements using inverse-variance fixed or random-effects methods and show forest plots. Check the input table, compatible units, assumptions, and included studies before interpreting a pooled estimate. If measurements do not match on variable, population, condition, or unit, keep them separate rather than forcing a synthesis.

Draft edits are not a substitute for a saved version. Use the explicit save action to persist an investigation snapshot; closing with unsaved changes prompts a warning. Export structured investigation JSON when another team needs the evidence model, and retain originals separately because a JSON export contains source references rather than file binaries.

Original files are stored in private object storage and historical snapshots remain available. Removing a source from a draft does not delete its original or prior saved versions. For transfers across accounts, reattach the original files as required and check the import preview before accepting the investigation.

A single original file is limited to 25 MB. PDF extraction considers at most 100 pages and 180,000 characters; OCR is limited to 100,000 characters and 1,000 passages per run. Spreadsheet extraction has sheet and row bounds. Large reviews should be split into related investigations and the application limits should be checked before import.

The default analyst uses local extractive lexical retrieval, not a generative or semantic model. Optional model analysis needs configuration and per-session consent; it validates citations and falls back locally on failure. Vann is research support, not an autonomous clinical or safety decision-maker. Organizational use still needs independent scientific and security review, backups, retention planning, accessibility review, and load testing.

Begin with a question that can be answered by a defined body of literature. Record the population or system, exposure or intervention, comparator, outcomes, study types, date range, and exclusions that matter. The protocol helps collaborators understand what the review is trying to include; it does not guarantee exhaustive retrieval from external databases.

Create an investigation for a coherent review, then add inclusion and exclusion criteria before screening begins. If the protocol changes, save the reason and identify which decisions need to be revisited. Keep a separate note for a hypothesis or interpretation so the eligibility rules do not become a retroactive justification for a preferred result.

Use discovery for Crossref and Europe PMC records when online access is available. Search result metadata may contain a title, authors, abstract, persistent identifier, date, and an access indicator. Inspect the originating provider and retrieval status before treating a source as available. Cached results can improve repeatability while still having an explicit retrieval timestamp.

If the source is open access, use the full-text route where supported and inspect the returned sections and page context. Otherwise, add the original document through an allowed import path. Confirm that duplicate records point to the same source identity rather than merging papers that share a title or author. Search coverage depends on the provider and its access terms.

Vann accepts supported PDF, DOCX, XLSX, text, and bibliographic inputs. Read the import preview: check citation fields, source identity, license notes, page count, sheet names, extraction status, and any warnings. The source library preserves the original attachment separately from the extracted text, which allows a later reviewer to inspect the file itself.

Digital text extraction, spreadsheet parsing, and OCR have explicit size and row limits. For scanned material, run the bundled English OCR on the page or image that needs it, then compare the extracted words and passage boundaries against the scan. Visual material such as plots, figure panels, and molecular structures must still be inspected in the source; Vann does not automatically interpret those images.

Review documents through a screening queue so each decision has a status and a reason. Use include, exclude, and uncertain states according to the protocol; an uncertain decision can mark a record for a second pass rather than forcing an early answer. Preserve the reviewer and the scope that applied when the record was assessed.

Screening is an administrative decision about a review's eligibility rules. It is separate from whether a paper's methods are strong, whether its claim is supported, or whether a measurement is usable in a synthesis. When criteria change, revisit affected records and keep the previous decision history rather than silently changing the saved review.

Select a source passage and create a claim with its text, source, page or section, and interpretation. Keep the exact quotation distinct from your paraphrase. Candidate claims produced by rules need human verification against the paper; the application can preserve a citation reference but cannot decide whether the author intended the same interpretation.

Connect experiments or study observations to the claims they address, then record support, contradiction, or uncertainty with a reason. Use annotations to explain terminology, caveats, or extraction problems. A reviewer should be able to move from a summary claim back to the original passage and understand which portion of the source was considered.

A quantitative record should include the variable, value, unit, sample context, condition, comparator, and source location. Validated CSV/TSV imports use rows shaped around `feature,sample,condition,value,unit`; exact identifiers and finite numeric values reduce ambiguity. Keep measurements and interpretation separate so a descriptive value is not mistaken for a causal conclusion.

Before pooling, compare outcome definition, units, population, condition, and uncertainty information. The inverse-variance fixed- or random-effects methods require compatible measurements and the relevant variance inputs. Inspect which studies entered the estimate, the forest plot, and the heterogeneity assumptions. If two values are not commensurate, retain a side-by-side evidence table instead of forcing them into a common summary.

The research map can display sources, authors, topics, claims, and entities with their recorded relationships. Use filters to focus on a source type or question and inspect the underlying records when a cluster or connection attracts attention. A graph can show what is linked in the workspace; it cannot infer that two papers are independent, that one author duplicated a cohort, or that a relationship is causal.

Contradiction candidates are a queue for human assessment. Compare the exact passages, study population, outcome, comparator, time point, and units before labeling evidence as conflicting. An apparent disagreement may reflect different contexts or definitions. Record the reason for a disposition so later readers can see how the initial signal was evaluated.

The default analyst performs local extractive lexical retrieval over selected investigation content. It can surface passages and structured records relevant to a query, but it is not generative semantic reasoning. Optional remote model analysis receives only the context selected for that request after explicit session consent; attachments can be opted in separately.

Returned citation IDs are checked against the investigation and malformed responses fall back to local behavior. That validation does not prove that a passage entails a generated sentence. Treat any model contribution as unreviewed interpretation until a person checks the cited text, corrects the wording, and records an accepted claim through the normal evidence workflow. Model access is not configured by default.

Draft changes require the explicit save action; closing with unsaved edits prompts a warning. Saved snapshots and audit records preserve prior states. Removing a source from a current draft does not delete the original or historical versions. A JSON investigation export contains structured evidence and file references rather than original binary attachments, so a cross-account handoff may require reattaching files.

The application limits file size, extraction pages and characters, spreadsheet rows, OCR passages, and snapshot size. Check the current in-app limits before building a large review and split a corpus into related investigations when needed. Back up the database and private object storage together. Vann supports literature review and evidence organization; it is not an autonomous clinical or safety decision-maker.