Eight features. One coherent workspace.
Each feature ships against the same primitive: the verified-quote stamp. Search a paper, open it, quote it, cite it, monitor it, validate it. The same provenance follows the quote from PubMed to your manuscript.
Verified Quotes.
Every quote is hashed against a page-anchored region in the source PDF before it ever reaches your document. No quote, no claim.
On retrieval, Docling extracts a page-anchored quote region. The agent's quote is matched verbatim to that region. Mismatches are rejected before the synthesis step. The hash and timestamp are persisted alongside the quote in your Markdown.
“the risk of worsening heart failure or death from cardiovascular causes was lower among those who received dapagliflozin”
Citation Chaining.
Traverse cited-by and cites two hops out. Rerank with MedCPT. Build a 47-node working graph from a single seed paper.
The Literature Reviewer issues graph.expand against the PubMed-derived citation network. Each node is rescored with MedCPT-Article against your question; nodes below threshold are dropped. The PRISMA log records every visited node and the inclusion/exclusion decision.
Multimodal PDF understanding.
Docling extracts text, figures, tables, and page-anchored regions in one pass. Quote regions are stored with byte offsets, not just page numbers.
docling.parse runs locally on the PDF. The output is a structured document tree: text spans, figure regions, table regions, all anchored to (page, paragraph, character) coordinates. Quotes resolve to specific regions; figures embed in your synthesis with their original captions.
Hybrid Retrieval.
BM25 first-pass over PubMed and your corpus. Dense rerank with MedCPT-Article. HyDE rewriting for short queries. The retrieval inspector shows the per-stage contribution to every result.
Three stages, observable. BM25 narrows the candidate pool; dense reranking applies a biomedical cross-encoder; HyDE generates a hypothetical answer to widen short, ambiguous queries. Each stage's contribution is logged to the run record so you can debug a missed paper.
Local-First Workspace.
Documents live as Markdown on disk. Embeddings, reranking, and optional LLM inference run on your hardware. Egress is opt-in, per call.
A standard project folder: manuscripts as Markdown, corpus as PDFs, embeddings in Qdrant on disk, retrieval models as gguf binaries. The Control Plane shows you which tools will leave the machine before each call. Diff your work in git like any code.
Bibliography Auto-synthesis.
Builds, deduplicates, and gap-analyses a working bibliography against the citation graph it inferred. Vancouver, APA, BibTeX, RIS export.
The Bibliography Generator agent maintains a structured set of references for the active document. Deduplication happens by DOI then PMID. Gap reports flag under-cited subtopics by year, author cluster, and embedding distance from the centroid of your set.
- Paediatric HFrEF: 0/30 papers
- SGLT2 in advanced CKD (eGFR < 30): 1/30 papers
Cross-Validation of Findings.
Locate a claim across a corpus; tag each match as direct, partial, indirect, or contradictory. The replication grid surfaces disagreement at a glance.
The Cross-Validator decomposes your claim into checkable sub-claims, runs each across the corpus, and emits a per-study verdict. The replication grid is exportable as a 1-page evidence card. Quotes are verbatim; verdicts are auditable.
Research Profile and Topic Monitoring.
Profile a researcher's prior work, build a topic neighbourhood, and subscribe to PubMed for new matches above a relevance threshold.
The Research Profiler clusters an author's output by MeSH descriptors and embedding distance. The Literature Monitor uses the resulting profile to set a daily PubMed subscription. Matches above threshold trigger a digest; matches below are logged but suppressed.
Compare Intelectra: vs Elicit · vs Consensus · vs Scite