Synthiz for researchers and PhD students

The problem: a literature growing faster than anyone reads it

A second-year PhD student typically has between 200 and 500 references in Zotero and has genuinely read about a third of them. For the rest there is a title, an abstract, and one sentence copied into a notes file. That is not negligence, it is arithmetic: the output of an active field exceeds what one person can read.

On top of that sits a layer almost nobody exploits: recorded seminars, keynotes, lab podcasts, streamed defences. Three hours in which a researcher explains in plain language what their paper says across twelve dense pages — sitting in a browser tab for six weeks.

Third problem: the trail. Six months on, you know a paper mentioned a threshold effect. Which one? You re-read three PDFs. Across a doctorate, that re-searching time produces approximate citations.

The workflow with Synthiz

1. Create one memoire per research question, not per broad topic: "Threshold effects in transfer policy" works better than "Public economics".
2. Load the corpus: article PDFs, preprint links, online chapters. The Chrome extension saves the open page in one click from the journal site, and captures YouTube transcripts when needed.
3. Answer "do I already know this?" without leaving the page. Mid-read, a hypothesis feels familiar: the extension's search panel queries your base from the current tab and shows the excerpt around the term. You know at once whether you already hold a paper on the point, and which one.
4. Add the spoken layer everyone skips: conference keynotes on YouTube, recorded seminars, research podcasts. Transcribed, they become searchable and citable at a precise passage.
5. Subscribe to a journal's RSS feed, a lab's video channel, a research centre's podcast: sync runs on its own and content arrives transcribed and summarised.
6. Read summaries to triage, not to replace reading: they tell you whether a paper is worth the two hours it demands.
7. Cross-analyse. The cross-analysis surfaces convergences, tensions and blind spots. A tension between two papers feeds your state of the art; a shared blind spot is a contribution lead.
8. Write thoughts anchored to sources — a methodological objection, an unexpected connection, a hypothesis. That is the raw material of your discussion section, dated and sourced.
9. Always verify before citing. Timestamped citations take you back to the exact passage; that is what goes into the manuscript, never the paraphrase.
10. Export to Markdown or PDF into your writing chain, LaTeX or word processor.

The problem the Zettelkasten does not solve: retrieval

The Zettelkasten and its software heirs — Obsidian, Roam, Logseq — are excellent thinking tools: link one note to another, let an unplanned connection emerge.

Their weakness lies elsewhere, and every third-year knows it: retrieval. You remember an argument about a threshold effect; finding which publication it came from means guessing the word you used in a note eighteen months ago, or having maintained a tag hierarchy that has since drifted. At citation time, approximation becomes a risk: you attribute to one author what another wrote.

Synthiz keeps concept linking and treats retrieval as a distinct problem. Search is hybrid: a semantic layer over embeddings, which finds a passage by meaning without requiring the paper's vocabulary, and a full-text lexical layer for author names, acronyms and numeric values.

The technical point is the refusal of a fixed threshold: relevance is decided by a document's distance from the distribution of its own query, plus an absolute floor. On a corpus of 948 embeddings, a query with 5 relevant documents and a query with no match peak at 0.09 apart — a fixed 0.65 threshold returned nothing to someone who did hold 5 sources. For citing: you find the paper without recalling its phrasing, and a query with no match returns zero rather than a neighbouring paper you would cite by default.

What it changes in practice

Task	Before	With Synthiz
Triaging 40 papers for a review	Abstract skimming, blind choices	Full summaries, informed triage
A two-hour recorded seminar	Never watched	Transcribed, summarised, searchable
"Who said that again?"	Re-reading several PDFs	Querying your own base
Spotting a disagreement in the literature	Emerges slowly, over months	Surfaced by the cross-analysis
Following a journal	Irregular manual checks	Automatically synced subscription

The real gain is not reading less: it is knowing sooner what deserves close reading, and no longer losing what has been read.

The honest limits

Synthiz does not manage your bibliography. No APA, Chicago or Vancouver styles, no BibTeX export, no word-processor plugin, no deduplication. Zotero, Mendeley or JabRef remain necessary: Synthiz works the content, they handle the reference.

It is not a qualitative analysis tool. No line-by-line coding, no coding frame, no inter-rater double coding: if your methodology rests on NVivo, Atlas.ti or MAXQDA, Synthiz does not replace them.

Summaries are not reading. Produced by language models, they structurally miss methodological subtleties: sample size, author-declared limitations, the exact statistical test used. A paper whose method you intend to discuss must be read — traceability is there to take you back to it.

No assessment of scientific quality. Synthiz does not distinguish a peer-reviewed journal from a preprint, or a meta-analysis from a study of 12 participants.

Transcription struggles with technical content. Formulae, mathematical notation, rare proper nouns, strong accents: errors are frequent. For a passage you intend to quote, go back to the recording.

No collaborative work. No corpus shared with your supervisor, no multi-person annotation: use is individual.

Nothing for research data. Datasets, analysis code, data management plans: out of scope. Synthiz handles text, audio and video.