Evidence-Centered Research — where Open Notebook is heading

[lfnovo]

This is the first initiative post under the model described in Open Contribution — The Bigger Picture: a direction, written down before it's a roadmap, so the community can shape it while it's still soft. It comes from months of maintainer analysis of where the product is straining and where the requests point. It is direction, not commitment — no dates, no promises. What we want from you is at the end.

The promise

Open Notebook exists to let you converse with your own content and get answers you can verify in the sources. Today it often produces a useful answer; it does not systematically deliver the two guarantees behind that promise:

1. find the information that's needed;
2. attribute it to the right passage — and make sure the passage actually supports the claim.

So the direction, in one sentence:

Every claim derived from your content should be verifiable in the evidence that supports it.

That sentence has consequences for search, chat, notes, podcasts, external sources, agents and the graph. This post is about those consequences.

What we heard

The requests below arrived as separate features. Read together, they are symptoms of a few platform decisions we haven't made yet:

• notebooks that don't fit the model's context — #1154, #473
• search that isn't wired into the conversation — #1036, #574, #427
• citations that depend on the model obeying the prompt — #294
• research agents and deep research — #292, #319, #973
• content that has to be ingested before it can be searched — web search #89, #342; external and mounted sources #291; Zotero #855
• knowledge that doesn't connect — knowledge graphs #248, timelines #122
• generated outputs asked for one by one — #203, images #566, re-runnable AI notes #1265, annotations #1215
• agents that want to use Open Notebook as their memory — #878

The common signal: people want Open Notebook to find better, connect better, cite better, and turn knowledge into new formats without losing where it came from. Treating that as a feature queue would produce a pile of disconnected systems. The alternative is fewer, shared foundations.

The shape: two systems and one contract

flowchart TD
    H[Human or agent] --> R[Search / Retrieval]
    R --> E[Evidence bundle]
    E --> A[Answer, note or artifact]
    A --> C[Verifiable citations and provenance]

Loading

1. Search and retrieval — finds material across sources, passages, notes, generated outputs, connected libraries and external sources. Used directly by you, or as a tool by agents.
2. Conversation and research — operates on that material at three scopes, which are the same capability at different sizes:

Scope	What it's for
Source	understand one piece of content deeply
Notebook	research and synthesize across a set
App	search globally across notebooks (the evolution of Ask)

Between them sits the evidence contract: a piece of evidence knows its content, its origin, whether it's original, derived or authored by you, where it sits in the source, what it relates to, and how it should be cited. Answers are built from evidence, and citations are validated by the system — not by trusting that the model wrote a correct ID in its markdown.

One structural note, because it changes where contributions land: the search–evidence–provenance core is being designed as a standalone, MIT-licensed open-source component that Open Notebook consumes, the same way it already consumes esperanto for models and content-core for extraction. Open Notebook stays the product — the notebook, the sources, the conversation, the community; the research core underneath becomes reusable by other tools, including agents.

The searchable universe is bigger than your Sources

A Source is content you decided to adopt and keep. The system should be able to search, use and cite content without turning it into a Source first:

Kind	Role	As evidence
Source	adopted, durable content	original evidence
External resource	web, Zotero, Obsidian, filesystem, other connectors	external evidence — adopt it if you want to keep it
Note	your own thinking	authored evidence, or a breadcrumb
Transformation / insight	derived interpretation	breadcrumb back to the original
Generated artifact	study guide, flashcards, podcast…	derived, searchable, traceable to its evidence

These stay distinct in the product. They are only normalized at the retrieval layer, which leads to the principle that runs through everything:

Any content can help find an answer; only its provenance decides how it may be used and cited.

A summary can guide a search; a note can point to related sources; but a factual citation should go back to the original passage whenever possible. Generated content becomes part of searchable memory, but its derived nature is never hidden — and every generated output keeps the evidence it was generated from, which is what makes "re-run this with today's sources" and "this is stale" possible.

Progressive agency: two modes, one core

Open Notebook has to work on a modest machine with a small local model and with frontier models and a budget. That's two execution profiles, chosen by you — not two products:

	Focused	Deep research
Goal	answer predictably, cheaply	investigate adaptively
Context	pre-selected by you, or retrieved once	starts compact, expands as needed
Tools	none or few	iterative search, graph, external sources
Models	small, local, no tool calling required	capable of tool use and long tasks
Cost	low, predictable	higher, variable

Both modes share the same scope, the same retrieval, the same evidence bundle and the same citation validation. Verifiability is not reserved for the powerful mode — what changes is how much effort the system is allowed to spend finding and confronting evidence. Focused mode is the baseline and works without embeddings or tool calling; a notebook chat gets an Auto context option that retrieves only what's relevant, server-side.

Reasoning inside or outside Open Notebook

The reasoning doesn't have to happen here. An agent in Claude, Codex or any MCP client can decide what to search and compose its own answer. That separates two things that shouldn't be confused: how much effort (focused vs. deep) and where the orchestration runs (our runtime vs. an external agent). MCP is not a third engine — it's an adapter over the same search-evidence-provenance core the internal modes use (#878). The first version would be local and read-only; exposing personal content remotely needs proper auth first, not the current password middleware.

The graph, in the right order

SurrealDB's graph is used today mostly for organization. The direction is progressive: first, relations we already know because a real action created them — belongs to, derived from, mentions, generated from, adopted from, supported by. That alone lets the system walk from a summary to its source, from an artifact to its evidence, from an adopted source to where it came from. Semantic extraction — entities, claims, contradictions, GraphRAG — comes later, as optional enrichment for global and multi-hop questions, never as a requirement for basic search or for small local installs.

Sequence, in broad strokes

No dates. Roughly:

• First: the retrieval and evidence contracts; structured citations in focused mode; scoped retrieval with hybrid search (text + vector, text-only fallback) and the Auto context option; fixing the current context plumbing so the server is the source of truth; a small evaluation set of real questions so improvements don't silently regress each other.
• Then: the shared research core extracted from today's separate runtimes into that standalone component; deep mode; local read-only MCP; external search and connected libraries; a common artifact infrastructure; the richer graph and remote surfaces.

Some of this is already moving: the UX foundation shipped as #1218 and #1220 so new interactions have a base to land on; #1226 fixed source chat context as a first plumbing step.

What we deliberately won't do

• Three different engines for Source, Notebook and App — or a fourth for MCP.
• Two evidence systems, one per mode.
• Require tool calling for simple conversations to work.
• Run a full agent on every question, or GraphRAG on every query.
• Extract a big knowledge graph before basic retrieval is proven.
• Treat citations as correct because the model wrote a valid ID.
• Treat derived outputs as original evidence.
• Ingest everything an agent finds on the web, or force-duplicate an Obsidian vault or Zotero library.
• Expose a remote MCP on today's password middleware.
• Put dozens of chunking and ranking knobs in the default UI.
• Rewrite the product before delivering value.

Where you come in

These are the questions where community input changes the design. Answer here, or in the linked thread if yours is one of them:

1. Citation granularity. Chunk + exact passage + source is the starting point. Which locators matter in your material — page, section heading, audio/video timestamp? What do you do when a passage repeats?
2. External sources first. Web search is the first family; SearXNG (self-hosted) and Brave (hosted) are the candidates. Which one unblocks you? For connected libraries (Zotero, Obsidian, filesystem — #291): live search on demand, or a local index that syncs?
3. "Cited" vs. "verified". The system can deterministically check that evidence exists, is in scope and matches the source. Checking that a passage semantically supports a claim may need another model call. Would you want that as an optional validator, an offline check, or a visible label — and at what cost?
4. Deep mode visibility. When the agent searches, opens sources and expands — what do you want to see? A timeline, a summary, only the budget and errors?
5. Naming. "Focused" vs. "Deep research"? Does the simple mode read as inferior? Better names welcome.
6. First artifacts. Which generated output would best prove the common infrastructure for you — re-runnable AI notes (#1265), flashcards, study guides, mind maps?
7. Scale. If you run notebooks with hundreds or thousands of sources (#1154), what breaks first today, and what would "good" look like?

How this becomes real

This thread is the deliberation. On October 1, 2026 we'll synthesize what came in. Whatever holds becomes an entry in VISION.md's Horizon and a decision record; the canonical threads linked above get a status refresh saying which part of the direction they feed; work items graduate the way #1270 describes — when someone is going to build them, not before.

Transparency, as in #1266: the underlying analysis and this write-up were AI-assisted; the direction, the cuts and every word here were reviewed and decided by the maintainer.qu

[lfnovo]

From the Discord thread on citations, two concrete asks that fit this direction:

1. Citations should point to the exact passage, not the whole source. Today the citation link opens the full source.
2. For PDFs, show the referenced page as an image rather than the extracted text. For graphical documents like rulebooks the extracted text is hard to read.

Also carried over from the "context management" thread there: make the difference between full content and insights obvious, make it easy to switch, and ideally have a smart default that needs no tweaking.

The Discord forum where these were discussed is now retired in favor of Discussions.