Extract from documents

Bootstrap a causal graph from your existing knowledge — proposals you review and accept.

Building a causal graph from scratch is a lot of typing. If you already have knowledge documents that describe the mechanisms in your domain, Nora can propose variables and edges automatically.

The extraction workflow

1. Pick a knowledge collection or specific documents.
2. Run extraction.
3. Review the proposals in a staged view.
4. Accept, reject, or edit each.
5. Accepted items land in your Draft graph.

Nothing changes in your live graph until you accept and publish.

Running extraction

1. Open the graph, click Extract from documents.
2. Choose the source:
   • A whole collection.
   • A tag-filtered slice.
   • Specific documents (multi-select).
3. Pick the extraction scope:
   • Variables only — find measurable quantities.
   • Edges only — assume variables exist, find relationships among them.
   • Both — end-to-end.
4. Run.

Extraction is asynchronous — you’ll get a notification when the staged proposals are ready to review. Typical time: 1-10 minutes depending on source size.

Reviewing proposals

The staging view groups proposals:

• New variables — with proposed name, type, unit, evidence quote.
• New edges — with proposed direction, sign, mechanism, evidence quotes.
• Merges — proposed unifications of similar variables (e.g., “customer_lifetime_value” and “clv” as the same variable).

Each proposal shows the source document, the quoted passage, and the confidence score. Click a proposal to see full context.

Accepting

Three actions per proposal:

• Accept — merges into Draft as-is.
• Edit & accept — modify (rename, change type, adjust sign) then accept.
• Reject — dismiss. Add a reason to help future extractions do better.

Bulk accept/reject the whole batch, or subset by confidence.

Confidence

Every proposal has a confidence score (0-1). High-confidence proposals often can be batch-accepted; low-confidence ones deserve individual review.

Filter the staging view: > 0.8 confidence for auto-approval-worthy, < 0.4 for likely-noise.

Evidence trail

Every accepted proposal carries a link back to its source document + quote. In the graph, right-click a variable or edge → View evidence to see what supported it. Useful for auditability and for revising when the source changes.

Re-running extraction

Re-running on the same source with the same settings is idempotent — Nora skips proposals it already made or that were previously rejected.

If a document was updated, extraction picks up on the new content and proposes changes to existing variables/edges.

Combining extraction and hand-authoring

Most graphs converge on a mix: extraction seeds ~60% of the structure, then experts hand-tune the important edges. That’s the intended workflow — don’t hesitate to override.

Extraction quality tips

• Cleaner sources → cleaner extraction. A well-structured doc extracts far better than a PDF scan.
• Domain-specific glossary helps. Attach a glossary document — extraction uses it to normalize variable names.
• Iterate. Extract, review, accept some, publish. Then extract again with a richer graph as context — later runs improve because they know your existing variables.