Positioning

Why this over NotebookLM?

Everything here helps you read faster. This is the only one that produces a synthesis a third party can check offline, without trusting us, without an AI, and without a network connection.

On the other columns. These products move fast and this is a best-understanding snapshot as of August 2026, not audited fact. The claims in this project’s own column were verified against the source; the rest is worth re-checking before you rely on it.

Feature by feature

yesyes
partialpartial
nono
not applicablenot applicable

Capability comparison across Research Assistant, NotebookLM, Google Scholar, Perplexity and Elicit
Capability	This	NotebookLM	Scholar	Perplexity	Elicit
Finds the literature for youNotebookLM is strongest on documents you already have.	yes	partial	yes	yes	yes
Grounded in your own documents	yes	yes	no	partial	partial
Citation resolves to a verbatim snippetNot just a link — the exact sentence the claim rests on.	yes	yes	no	no	partial
Says when it cannot verify a citationA failed check renders a ⚠ chip rather than rendering clean.	yes	no	not applicable	no	no
Standalone offline verifierNo AI, no network, no account, no trust in us.	yes	no	no	no	no
Tamper-evident exportEditing the report after approval breaks the hash chain.	yes	no	no	no	no
You approve the plan before it spends	yes	no	not applicable	no	no
You approve the draft before it is final	yes	no	not applicable	no	no
Evidence only from your own documents, retrieved on your machineCorpus-only mode: evidence from your uploads only, web search off, local indexing model required. The question and excerpts still go to the provider each research model uses, a cloud provider unless you choose local models; the Offline switch refuses to start unless every model is at an address on your computer.	yes	no	no	no	no
Runs on local models	yes	no	not applicable	no	no
A different model per agent	yes	no	not applicable	no	no
Self-hostable	yes	no	no	no	no
Your own API keys	yes	no	not applicable	no	no
Cost visible before it is spent	yes	not applicable	not applicable	no	no
Polish, scale, ecosystemSmall and new. This is a real gap, not a rounding error.	no	yes	yes	yes	yes
Audio overviewsNotebookLM's are genuinely good and there is no answer to them here.	no	yes	no	no	no
Zero setupDocker or a desktop app, plus your own keys.	no	yes	yes	yes	yes

Finds the literature for you

NotebookLM is strongest on documents you already have.

This: yes
NotebookLM: partial
Scholar: yes
Perplexity: yes
Elicit: yes

Grounded in your own documents

This: yes
NotebookLM: yes
Scholar: no
Perplexity: partial
Elicit: partial

Citation resolves to a verbatim snippet

Not just a link — the exact sentence the claim rests on.

This: yes
NotebookLM: yes
Scholar: no
Perplexity: no
Elicit: partial

Says when it cannot verify a citation

A failed check renders a ⚠ chip rather than rendering clean.

This: yes
NotebookLM: no
Scholar: not applicable
Perplexity: no
Elicit: no

Standalone offline verifier

No AI, no network, no account, no trust in us.

This: yes
NotebookLM: no
Scholar: no
Perplexity: no
Elicit: no

Tamper-evident export

Editing the report after approval breaks the hash chain.

This: yes
NotebookLM: no
Scholar: no
Perplexity: no
Elicit: no

You approve the plan before it spends

This: yes
NotebookLM: no
Scholar: not applicable
Perplexity: no
Elicit: no

You approve the draft before it is final

This: yes
NotebookLM: no
Scholar: not applicable
Perplexity: no
Elicit: no

Evidence only from your own documents, retrieved on your machine

Corpus-only mode: evidence from your uploads only, web search off, local indexing model required. The question and excerpts still go to the provider each research model uses, a cloud provider unless you choose local models; the Offline switch refuses to start unless every model is at an address on your computer.

This: yes
NotebookLM: no
Scholar: no
Perplexity: no
Elicit: no

Runs on local models

This: yes
NotebookLM: no
Scholar: not applicable
Perplexity: no
Elicit: no

A different model per agent

This: yes
NotebookLM: no
Scholar: not applicable
Perplexity: no
Elicit: no

Self-hostable

This: yes
NotebookLM: no
Scholar: no
Perplexity: no
Elicit: no

Your own API keys

This: yes
NotebookLM: no
Scholar: not applicable
Perplexity: no
Elicit: no

Cost visible before it is spent

This: yes
NotebookLM: not applicable
Scholar: not applicable
Perplexity: no
Elicit: no

Polish, scale, ecosystem

Small and new. This is a real gap, not a rounding error.

This: no
NotebookLM: yes
Scholar: yes
Perplexity: yes
Elicit: yes

Audio overviews

NotebookLM's are genuinely good and there is no answer to them here.

This: no
NotebookLM: yes
Scholar: no
Perplexity: no
Elicit: no

Zero setup

Docker or a desktop app, plus your own keys.

This: no
NotebookLM: yes
Scholar: yes
Perplexity: yes
Elicit: yes

The three claims nobody else makes — and how they are built

01Every [n] is falsifiable

How The executor submits evidence carrying a verbatim snippet. The synthesizer is shown only that snippet — never the executor's paraphrase, because a model shown both writes from the paraphrase and drifts past the quotable text. A separate pass then checks each cited sentence against its snippet, and a failure is flagged rather than hidden.

Why it matters NotebookLM links you to a passage. This tells you when the link does not hold up — the difference between “here is where I got it” and “I checked, and this one is weak”.

research_engine/graph.py

02A third party can verify without trusting you

How The .bundle.json export carries the report, evidence, sources, contradictions, the models actually dialled, cost, and the approval chain — with the report hashed against the draft a human approved, and a bundle hash covering every field except itself. A standalone verifier checks all of it offline.

Why it matters An advisor, reviewer or journal can confirm the synthesis was not edited after approval without an account, an internet connection, or trusting the vendor. For contested work that is a category difference, not a feature.

research_engine/verify_bundle.py

03Your unpublished manuscript can stay on your machine

How Corpus mode delegates retrieval exclusively to the local store, refuses every non-corpus URL, and fails closed — an embedder that does not declare itself local is treated as remote and rejected. That keeps the evidence local, not the run: the excerpts still reach the planner, executor, critic and synthesizer, so the manuscript stays local only when all four run on local models. The Offline switch checks that each model's address is on this computer; a local router that forwards to a cloud provider passes it.

Why it matters A cloud product structurally cannot offer this. With embargoed data, patient records or an unpublished draft, “upload it to Google” is not a decision you are allowed to make.

backend/tests/dataflow/test_corpus_egress.py

Who should actually use this

PhD student writing a literature reviewUse thisThe citation chain is the deliverable, and the design gate makes the structure yours.
Researcher with confidential or embargoed dataUse thisThe only option here whose evidence and models can all run on your own machine.
Anyone whose institution forbids cloud uploadUse thisSelf-hosted, on your keys or your own GPU.
Someone whose output has to survive scrutinyUse thisThe offline verifier exists for exactly this.
Reading five PDFs you already haveUse NotebookLMBetter at it, free, no setup, and the audio overviews are excellent.
Just trying to find papersUse ScholarIt is search. This is not.
Want a quick answer with linksUse PerplexityFaster, and there is nothing to install.
Non-technical, no Docker, no API keysNot yetThe setup cost is real and it is the loudest weakness here.

Where it honestly loses

A comparison written by the thing being compared is worth nothing unless it is willing to lose. These are the reasons to pick something else.

Setup is a real barrier. NotebookLM is a URL; this is Docker or a desktop app plus API keys.
No audio overviews, no mobile app, no collaboration. NotebookLM's polish is far ahead.
Small and new — no ecosystem, no team features, one maintainer's budget.
The headline metric rests on one run: citation support is 96.4% across ten fixed questions, graded by an independent model rather than the pipeline's own. Ten questions on one model routing is a small sample, and support means a claim matches the evidence it cites, not that the claim is true.
Quality tracks the models you route it at. Point it at a weak local model and you get weak research — the verification machinery will report that honestly rather than paper over it.

Read further

The documentation describes what is built, not what is planned — including the agent design, the local-first architecture, and the security model.

Documentation →Download →