notebooklm-deep-research

Use NotebookLM's research-agent capability for questions that need depth beyond a chat query — multi-step reasoning, web search, Drive scan. Returns a structured report stored as a wiki node so it stays linkable.

Pre-flight checks

Wiki/Libraries/{library_slug}.md exists.
NotebookLM session valid. Phase 1 fallback: scaffold an empty research note with status: pending-installation and surface to Adam.

Procedure

1. Fire the research job (Phase 2+)

Use notebooklm-py's research-agent with mode={fast|deep} and auto_import={bool}.
Capture report_id. The library may take 30s–several minutes (deep mode).

2. Wait & retrieve

Poll for completion (with a sane timeout — deep mode can be 5+ min).
On completion, fetch the structured report (markdown).

3. Write the wiki node

Path: Wiki/Notes/research-{YYYY-MM-DD}-{question-slug}.md

Use the wiki-node template. Frontmatter:

type: wiki-node
node_kind: concept
node_subkind: research-report
created: {today}
last_synthesized: {today}
sources:
  - "<a class="wikilink wikilink-broken" href="#">Wiki/Libraries/{library_slug}</a>"
links: []                            # filled in step 4
status: synthesized
confidence: medium
tags: [research, source/notebooklm]
notebooklm_report_id: {report_id}
research_question: "{question}"
research_mode: {mode}

Body: NotebookLM's report markdown, lightly cleaned.

4. Connect to the graph

Scan the report for entities, decisions, brands, people. For each match against existing wiki nodes, add a <a class="wikilink wikilink-broken" href="#">wikilink</a> to links: frontmatter and inline in the body.
If the question is decision-shaped, optionally cross-link to Departments/Decisions/Memos/ if a relevant memo exists.

5. If `auto_import` was true

Run notebooklm-sync for library_slug so any new sources NLM imported get reconciled into the wiki Sources table.

6. Append Log entry

input: library={library_slug}, question="{question[:80]}...", mode={mode}, auto_import={auto_import}
output: report_path={path}, report_id={report_id}
artifacts: <a class="wikilink wikilink-broken" href="#">Wiki/Notes/research-{date}-{slug}</a>, <a class="wikilink wikilink-broken" href="#">Wiki/Libraries/{library_slug}</a>

When to use this vs `notebooklm-query`

Use case	Use
Quick fact lookup, single-paragraph answer	`notebooklm-query`
Multi-source synthesis, "explain this topic comprehensively"	`notebooklm-deep-research`
Decision needs grounding (Steelman/Devils-Advocate phase)	start with `notebooklm-query`; escalate to deep-research only if the dialectic needs more depth
Building out a new wiki concept hub	`notebooklm-deep-research` then synthesize the report into atomic `Wiki/Concepts/` nodes

Hard rules

Deep research jobs are expensive (rate-limit + time). Don't fire reflexively — confirm the question warrants the depth.
Never auto_import: true without Adam's explicit confirmation — it can pull in surprising sources from web search.
Always store the report as a wiki node (not just in conversation) — reports take effort; preserve them.

Failure modes

Job times out — log as failed; let Adam retry.
Job returns empty (rare; happens with very narrow library + broad question) — prompt Adam to widen the library or sharpen the question.
Auto-imported source is gated/paywalled — NLM will report the failure inline; pass through to the report and flag.