Research

"What would you like to research? You can ask a direct question, explore a topic, compare authors, search for quotes, or browse your sources."

Accept any of these forms:

• A direct question: "What does Foucault say about power in Discipline and Punish?"
• A topic exploration: "Show me everything I have on settler colonialism"
• A comparison: "Compare how Author A and Author B treat the concept of sovereignty"
• A source lookup: "What sources do I have about Ottoman Palestine?"
• A quote search: "Find the exact quote where Ben-Gurion talks about..."

Parallel Research via Subagents

Speed is critical. Spawn subagents in parallel to query all enabled tools simultaneously. Do NOT query tools sequentially.

Determine which subagents to launch

Based on the question and enabled tools, spawn ALL applicable agents at the same time:

Use the Agent tool to spawn subagents in parallel — call the Agent tool multiple times in a single response:

1. RAG subagent: Pass as prompt the question, questionType, sourceIds. This subagent runs multi-mode RAG queries (mix + local + bypass as needed).

2. Candlekeep subagent (if tools.candlekeep.enabled): Pass as prompt the question, questionType, sourceIds. This subagent searches and reads Candlekeep documents directly.

3. NotebookLM subagent (if tools.notebooklm.enabled): Query existing notebooks or create a temporary research notebook. Use the notebook_query MCP tool for AI-powered Q&A against indexed sources. NotebookLM excels at synthesizing across multiple sources and identifying connections that keyword search may miss.

Each subagent runs independently and returns its findings.

RAG Subagent Instructions

The RAG subagent should:

1. Classify the question type and select RAG modes:

2. Run all queries with include_context: true, top_k: 30, rerank_top_k: 10, enable_rerank: true
3. Return organized passages with source attribution from the context field
4. Never cite the answer field — only context

Candlekeep Subagent Instructions

The Candlekeep subagent should:

1. Search across all sources: ck items search "QUERY"
2. Read relevant documents for full passages: ck items read DOC_ID
3. Verify page numbers when exact citations are needed: ck items read "DOC_ID:PAGE-PAGE"
4. Return full text passages with document ID, title, and page numbers

NotebookLM Subagent Instructions

The NotebookLM subagent should:

1. List existing notebooks: Use notebook_list MCP tool to find relevant notebooks
2. Query notebooks: Use notebook_query MCP tool with the research question for AI-powered Q&A
3. Cross-reference: NotebookLM answers are AI-synthesized — always cross-reference key claims against Candlekeep/RAG results for citation accuracy
4. Return findings with notebook name and any source references NotebookLM provides

Important: NotebookLM answers are supplementary context. Never cite NotebookLM output as a primary source — always verify claims against Candlekeep or RAG results.

Merge & Present Findings

Once all subagents return, merge results by deduplicating and cross-referencing:

1. If both RAG and Candlekeep found the same passage, use Candlekeep's version (more precise page numbers)
2. If RAG found something Candlekeep didn't, include it with a note that page numbers come from RAG
3. If Candlekeep found full text that RAG summarized, prefer the full text

For factual questions:

Answer: [Direct answer grounded in sources]

Sources:

• [Author], [Work], p. [page] — "[relevant quote]" (via Candlekeep/RAG)
• [Author], [Work], p. [page] — "[relevant quote]" (via RAG)

Context: [Brief explanation of how these sources relate]

For topic explorations:

What your sources say about [topic]:

1. [Theme/aspect] [Summary with citations]

2. [Theme/aspect] [Summary with citations]

Available sources on this topic:

• [list of relevant documents from Candlekeep/profile]

Gaps: [What your sources don't cover]

For comparisons:

[Author A] vs. [Author B] on [concept]:

Aspect	Author A	Author B
...	... (with citations)	... (with citations)

Key differences: ... Points of agreement: ...

For quote searches:

Found in [Author], [Work], p. [page]:

"[exact quote]"

Context: [surrounding text / chapter / argument] Verified: ✓ Candlekeep page confirmed / ⚠ RAG only (page may need verification)

Cognetivy Logging (if enabled)

Log the research session:

echo '{"subject": "RESEARCH_TOPIC", "type": "research"}' > /tmp/aw-research-input.json
cognetivy run start --workflow wf_academic_writer --input /tmp/aw-research-input.json

Log query results:

echo '{"type":"step_completed","nodeId":"research","ragQueries":N,"candlekeepReads":N,"totalPassages":N,"toolsUsed":["rag","candlekeep"]}' | cognetivy event append --run RUN_ID

Follow-up

After presenting findings, ask:

"Would you like me to:

• Dig deeper into any of these sources?
• Search for related topics?
• Find exact page numbers for any of the citations?
• Save these findings for use in an article?"

Critical Rules

• NEVER fabricate citations. Only cite what RAG or Candlekeep returns.
• NEVER invent page numbers. Use bypass mode + ck items read to verify.
• If a question can't be answered from available sources, say so clearly: "Your indexed sources don't contain information about this. You may need to add more sources via Candlekeep."
• Always show which tool provided each piece of information.
• NotebookLM is supplementary — its AI-synthesized answers must be verified against Candlekeep/RAG before citing.
• If a tool is disabled, skip it and note: "I searched [available tools]. Enable [disabled tool] with /academic-writer:update-tools for broader coverage."