9 June 2026Biomedical research workspace · grounded in PubMed

ChatGPT guesses.
Intelectra reads the paper.

The biomedical AI workspace built on verifiable evidence, not autocomplete. Five specialist agents that search PubMed, read the full text, and quote the source, never invented.

5
Specialist agents · one workspace
0
Hallucinated citations · ever
35M+
PubMed abstracts · indexed, refreshed nightly
McMurray et al. · N Engl J Med · 2019verbatim · verified

Heart failure with reduced ejection fraction (HFrEF) carries a high risk of progression. In the DAPA-HF trial, we evaluated whether sodium–glucose cotransporter 2 (SGLT2) inhibition with dapagliflozin would reduce that risk, regardless of diabetes status. Among patients with heart failure and a reduced ejection fraction, the risk of worsening heart failure or death from cardiovascular causes was lower among those who received dapagliflozin than among those who received placebo, regardless of the presence or absence of diabetes. These benefits were consistent across prespecified subgroups and emerged early in follow-up.

citeDAPA-HF · NEJM 2019 · PMID 31535829
locusp.1999 · §results ¶1
checkverbatim matchsource paper
Grounded in
PubMed (NCBI)MeSHEuropePMCPeer-reviewed full textHybrid retrievalFrontier LLMs
§ 01The problem

Three things generalist AI gets wrong, every time.

Three failure modes show up in every research workflow built on a generalist chatbot. Intelectra answers each one structurally, not as a feature flag.

FAIL · 01

Generalist AI invents the citations.

Ask ChatGPT for a reference and it will write you a PMID that doesn't exist, an author who never co-published, a journal volume that was never printed. The model isn't lying; it's autocompleting. Intelectra refuses to emit a quote unless it matches a verbatim region of a real paper.

“Smith et al. (2023) showed efficacy in 78% of patients (PMID 91428557).”
McMurray JJV et al. (2019). DAPA-HF. NEJM. PMID 31535829. Verbatim quote, p.1999 ¶1.
FAIL · 02

Chatbots only read the abstract. We read the paper.

A 250-word abstract isn't the evidence. The Methods section is. The results table is. The footnote where the trial reports a 4-week follow-up window is. Intelectra retrieves the full text and quotes verbatim from the actual page, not from a summary an indexer wrote.

Generalist AI seesabstract · ~250 words · trial conclusionIntelectra readsfull text · methods · results · footnotes
FAIL · 03

Your manuscript shouldn't train someone else's model.

Pasting unpublished work into a hosted chat sends your draft, your figures, and your hypotheses to a third-party provider that may train on it, log it, or subpoena it. Intelectra runs in three modes: open, hybrid, or fully self-hosted. You decide where your evidence lives.

open third-party LLM · Intelectra-hosted services
hybrid your LLM · Intelectra-hosted services
closed your LLM · your servers · zero external egress
You draw the boundary. See the modes →
§ 02How it works

Understand. Search. Retrieve. Verify. Cite.

Five named stages. The same pipeline that powers every demo on the public sandbox runs inside the desktop workspace, on your corpus, with your own boundary line drawn around it.

01

Understand

We decompose your question, paragraph, or hypothesis into atomic claims a paper can actually answer.

Decompose · Reformulate
02

Search

We query PubMed with MeSH expansion and walk the citation graph two hops out: references and cited-by.

PubMed · MeSH · Citation chaining
03

Retrieve

Hybrid retrieval combines exact term matching with biomedical semantic embeddings, then re-ranks the top candidates.

Hybrid retrieval · Biomedical reranker
04

Verify

Every quote is checked against the source paper, verbatim or semantically. Anything that doesn't survive verification is dropped before you ever see it.

Verbatim check · Factored verification
05

Cite

The answer is assembled from verified quotes only: Vancouver-style citations, no claim without a source.

Grounded synthesis · No quote, no claim
§ 02bLive demos

Five live demos. No signup. No mocks.

Each demo runs the same retrieval and verification pipeline against PubMed. Click a card to run it in your browser.

Run a real example
§ 03The five

Five specialists. One workspace.

The desktop workspace ships with five specialist agents that collaborate with you on the two jobs every biomedical researcher hates doing manually: building a bibliography, and writing a review of the literature.

§ 04Three modes

Three modes. You draw the boundary line.

Intelectra runs in three deployment modes. The retrieval and verification pipeline is identical across all three. What changes is where the LLM runs, and where the services that orchestrate it live. Same product, three privacy postures.

MODE · 01OpenLLMTHIRD-PARTYAgent orchestrationINTELECTRA · EUROPERAG + paper databaseINTELECTRA · EUROPEPubMed retrieverINTELECTRA · EUROPEBEST FORPublic research · fastest setupMODE · 02HybridLLMYOUR GPUAgent orchestrationINTELECTRA · EUROPERAG + paper databaseINTELECTRA · EUROPEPubMed retrieverINTELECTRA · EUROPEBEST FORNDA-sensitive prompts · neutral retrievalMODE · 03ClosedLLMYOUR GPUAgent orchestrationYOUR SERVERSRAG + paper databaseYOUR SERVERSPubMed retrieverYOUR SERVERSBEST FORAir-gapped labs · pharma R&D · clinical
See the modes in detail
§ 05Comparison

Where we win, and against whom.

Generalist chatbots and biomedical search tools, side by side. Categories we built the product around, categories we still treat as table-stakes. The columns aren't a leaderboard; they're a clarity exercise.

IntelectraChatGPTGeminiPerplexityElicitOpenEvidence
Refuses to invent citations
Reads the full paper, not just the abstract
Native PubMed + MeSH retrieval
Citation chaining (references + cited-by)
Verbatim quote verification
Bibliography as persistent state
Cross-paper claim triangulation
Self-hostable for unpublished work

Stop trusting AI that bluffs. Cite from the paper.

Run a real example on the public sandbox, or claim a spot on the desktop build. Researchers like you are already shaping what ships first.