Five specialists. One workspace.

A structured bibliography plus a thematic Markdown draft where every claim is backed by a verbatim quote from a real paper.

1. Seed: you pick 3–5 anchor papers; we walk references and cited-by recursively to 40–60 papers (60–100 for a thesis).
2. Verify: every quote we extract runs through a three-method verbatim check; nothing that fails verification reaches your draft.
3. Synthesize: we cluster the verified quotes into themes, write a gap report, and draft the review with inline citations to every claim.

The review draft is written by your local LLM; citation chasing and full-text retrieval still run against the shared biomedical corpus. Best balance for most labs.

25–45 references organised by the role they play in your introduction, methods, and discussion. Every citation is backed by a verified quote.

1. Decompose: we read your abstract or project aims and identify the 6–14 claims that need a citation.
2. Fill: for each claim, we search PubMed, propose 3–5 candidates, and validate each one against the source paper.
3. Balance: we check the funnel is complete (no missing methodological references, no over-cited subtopic), then export to BibTeX, RIS, APA, or Vancouver.

Your abstract is decomposed by your local LLM; citation enrichment and verbatim grounding still query PubMed. Best balance when your draft is sensitive but the literature is not.

A one-page evidence card: direct support, partial support, indirect contradiction, direct contradiction, each with the verbatim quote and the methodology behind it.

1. Locate: we hunt across PubMed and your indexed corpus for every paper that touches your claim, regardless of conclusion.
2. Classify: we tag each match by stance and methodology (RCT, observational, meta-analysis, mechanistic) and weight accordingly.
3. Surface: we report disagreement and replication explicitly. We never average across studies; we report the studies.

Verification runs on your local LLM; retrieval still hits PubMed. Best balance for most labs.

A morning digest of new biomedical publications matching your saved MeSH descriptors, each scored for relevance to the corpus you actually care about.

1. Pick: we help you choose the right MeSH descriptors (the controlled vocabulary PubMed uses) instead of fragile keyword strings.
2. Watch: the nightly PubMed delta inspects the freshly-ingested papers and emits alerts for the ones matching your tags.
3. Score: each alert is ranked against the rest of your corpus, so the most relevant new papers surface first.

Relevance scoring is done by your local LLM against your private corpus; the PubMed crawl that surfaces new papers still happens on our side. Best balance when the watchlist is sensitive but PubMed isn't.

A concise answer with inline citations to real papers, and an honest 'no good evidence' when the literature doesn't support a claim.

1. Decompose: we break your question into searchable sub-questions a paper can actually answer.
2. Retrieve: we hit PubMed and your indexed corpus, fetch the full text when needed, and rank the top candidates.
3. Cite: every claim in the answer is paired with the verbatim quote that supports it; unsupported claims are flagged, not invented.

The chat LLM runs on your local model; retrieval and verbatim grounding still query PubMed. Best balance when your questions are sensitive but the literature is public.