Why Put a Boundary Around AI Research?

Open Notebook is a self-hosted research workspace built around a deceptively practical question: which parts of your research should an AI system be allowed to see, use, and reshape? This opening focuses on the boundary between a person's source material and the model that helps interpret it. Understanding that boundary gives the reader a way to judge Open Notebook's value, rather than treating it as just another chat window with document uploads.

The distinction matters because “AI-assisted research” hides several different activities. A researcher may want to preserve an original paper, search across a collection, discuss two sources, extract the same fields from many documents, save a conclusion, or listen to a spoken version while commuting. Those activities do not require the same context, the same model, or the same degree of trust. Open Notebook's design is an attempt to keep them connected without making them identical.

A research workspace, not a blank chat window

Open Notebook describes itself as a privacy-focused, open-source alternative to Google's NotebookLM. That comparison explains its starting point, but not its whole identity. The project is a research assistant for managing and understanding content; it connects to different AI providers; it can be operated through a web interface or an API; and it can be extended through standard interfaces. Its own vision also draws a boundary around what it is not. It is not intended to be a general-purpose chatbot, a document editor, or a replacement for every tool in a researcher's workflow.

That boundary changes how the product should be understood. A general chatbot begins with a conversation and treats attached material as supporting context. Open Notebook begins with a body of research that the user chooses to organize. The conversation is one way to work with that body of research, not the container that defines it.

The basic unit is a notebook: a scoped workspace for a project or question. Inside it are sources, such as PDFs, web pages, text, recordings, or videos. Sources are the incoming material. They may be extracted, divided into searchable pieces, and represented for semantic search, but their role remains to provide the material being studied.

A second layer consists of insights and notes. An insight is a generated result associated with a source, often produced by applying a transformation such as a summary or a structured extraction. A note may be written by the user, saved from a conversation, generated from an answer, or built from an insight. Notes are where interpretation becomes useful for later work: they can be edited, tagged, searched, and reused. A chat session is another persistent artifact, preserving a line of questioning rather than merely returning a disposable answer.

This arrangement is more important than the names suggest. It separates at least three things that are easy to confuse:

  1. What the source contains.
  2. What the system derived from the source.
  3. What the researcher decided to retain as working knowledge.

If a generated note says that a report supports a particular conclusion, the note is convenient, but it is not the report. The useful question is still: which source passage supports the conclusion, and did the generated interpretation preserve its qualifications? Open Notebook's model makes that return to the source possible. It does not make the return unnecessary.

A boundary is made of relationships

It is tempting to think of a notebook as a folder. A folder answers, “Where is this file?” A notebook answers a more consequential question: “Which research context does this material belong to, and which operations are allowed to draw on it?” That is why the project's current data model uses relationships among records rather than relying only on filenames or labels.

A source is connected to a notebook through a relationship, and a note is connected through a corresponding artifact relationship. A chat session is related to the notebook whose research it explores. Source insights remain tied to their parent source. These links give the system something concrete to filter when a user chooses a notebook scope for search or Ask. They also let the application handle deletion more carefully: a source linked to more than one notebook is not the same as a source used only by the notebook being removed.

erDiagram NOTEBOOK ||--o{ REFERENCE : scopes SOURCE ||--o{ REFERENCE : joins SOURCE ||--o{ SOURCE_INSIGHT : produces NOTEBOOK ||--o{ ARTIFACT : contains NOTE ||--o{ ARTIFACT : captures NOTEBOOK ||--o{ CHAT_SESSION : hosts

This relationship model exposes an important qualification. Some conceptual guidance describes a source as belonging to exactly one notebook and as remaining unchanged once added. The current implementation is less rigid. A source can have multiple notebook relationships, and processed content can be reprocessed; its title and topics can also be edited. The practical principle is not that a source is physically untouchable. It is that source material should remain distinguishable from the interpretations built on top of it, and that any change or reprocessing should be treated as a change in the evidence being used.

Notebook scope is therefore a retrieval rule, not a guarantee that the database contains nothing else. Search and Ask can operate across the whole knowledge base when no scope is supplied, or across the sources and notes linked to selected notebooks when a scope is supplied. A result that does not appear under a selected notebook may still exist elsewhere. Conversely, a source shared across notebooks can participate wherever the relationship and the chosen scope allow it.

That distinction matters in ordinary research. Imagine one notebook for public literature and another for an internal product investigation. The same public source might be useful in both. An internal design document may belong only to the second. A global search can mix them; a scoped search can limit the field. The boundary is effective only when the user chooses the scope that matches the question.

The same logic applies to notes. A note is not merely a shorter source. It is a mutable statement that may combine several sources, a user's judgment, and a model's wording. Because notes are searchable and can be supplied as context later, a careless note can become a second-order source: future answers may rely on the note rather than returning to the original document. The benefit of accumulation comes with a responsibility to preserve links and mark the difference between evidence and synthesis.

The question Open Notebook lets you control

The project’s central promise can be stated simply: your research should remain under your control while AI remains useful. “Control,” however, has several layers.

First, there is control over where the records live. A self-hosted deployment stores the application's data and database within infrastructure chosen by the operator. This is materially different from placing an entire research collection in a provider's hosted workspace. It also means the operator is responsible for the database, persistent volumes, backups, network exposure, and updates.

Second, there is control over which provider performs which task. Open Notebook supports a provider registry and separate model records for language, embedding, speech-to-text, and text-to-speech work. A user can connect a hosted model, a local Ollama or compatible endpoint, or a mixture of both. That flexibility is a design choice against automatic dependence on one vendor, but it is not a promise that all providers offer the same capabilities or produce equivalent results.

Third, there is control over what reaches a model in a particular interaction. The notebook Chat path can be built from selected sources and notes. Items can be excluded, supplied in a shorter insight-oriented form, or supplied with longer content. Source Chat handles one source under its own context budget. Ask takes a different route, searching for relevant material before asking a model to synthesize it. These are different ways of drawing the line between the stored research collection and the prompt sent to a model.

That line is also a cost and quality decision. Supplying more content may give a model more opportunities to notice a qualification, but it can increase token use and distract the response. Supplying less content can be cheaper and more focused, but it can omit the passage that changes the meaning. Automatic retrieval can find material the user did not think to select, but it can also miss relevant context. Manual selection is transparent, but it depends on the user's judgment.

Finally, there is control over what becomes shareable output. A researcher may keep a source in the notebook, turn an analysis into a note, export a result, or generate a podcast. Each step changes the form of the material. None of them changes the underlying obligation to distinguish the original evidence from the system's processing of it.

This is why self-hosting should not be equated with being offline. If the database is local but an external language model receives the source text, the workflow is hybrid. If a local model writes the script but a cloud text-to-speech provider voices it, some generated content still leaves the deployment. If a web page is fetched through a hosted extraction service, that fetch introduces another boundary. A genuinely local workflow requires examining every relevant stage, including extraction, embeddings, language generation, transcription, and audio generation.

The reverse is also worth keeping in view: a hosted provider is not automatically incompatible with responsible research. Some material may be public, some tasks may benefit from a stronger hosted model, and a user may decide that the trade-off is acceptable. The important point is that the decision should be visible. Open Notebook's value lies less in declaring one universal answer than in giving the user places where the answer can be chosen.

That choice does not eliminate uncertainty. Search rankings indicate relevance, not truth. A generated insight can be fluent and wrong. A citation can point to a related passage without proving that the passage supports the exact claim. The appropriate habit is therefore layered: use the notebook to scope the material, use processing and retrieval to find what matters, use notes to preserve working understanding, and return to the source when a claim matters.

The rest of this series follows that boundary through the system. The next question is the most concrete one: what, exactly, is a notebook made of, and how do sources, insights, notes, and conversations remain separate enough to be trusted while connected enough to support real research? Open Notebook's answer begins with its container model—not as a filing cabinet, but as the structure that determines what can be found, remembered, revised, and shown to AI.

References