Literature Research

Ocupación

Categorías

Base de Conocimientos

Do NOT use this skill for extracting details from the author's own papers (001)). Use paper_analysis instead. Do NOT use this skill for deep extraction from review articles (002)). Use review_analysis instead. This skill identifies relevant reviews; review_analysis extracts from them.

Prerequisites

1. Run source_registry first to obtain the Review Articles (002)) source IDs.
2. This skill uses a two-phase query strategy (see below).

Two-Phase Query Strategy

Research follows two phases: start narrow with curated reviews, then widen to all sources for gap-filling.

Phase 1: Reviews Only (Big Picture)

Query only Review Articles (002) source IDs from source_registry) to establish the big picture, consensus definitions, and structural context.

mcp_notebooklm_notebook_query(
    notebook_id="1b7df790-7858-4fc8-879c-39f41238c4ae",
    query="<your question>",
    source_ids=<review_002_ids>  # only the ~8 review source IDs
)

Why reviews first? Reviews are curated, authoritative, and provide the "big picture." Starting here ensures the foundation is built on consensus before drilling into specifics.

Phase 2: All Sources (Gap-Filling & Specifics)

For follow-up questions, clarifications, or finding specific papers that reviews mentioned but didn't detail — query all sources by simply omitting source_ids (defaults to the entire notebook).

mcp_notebooklm_notebook_query(
    notebook_id="1b7df790-7858-4fc8-879c-39f41238c4ae",
    query="<your follow-up or specific question>",
    conversation_id=<previous_conversation_id>  # maintains context from Phase 1
)

Why omit source_ids? With 170+ general references, passing them individually is impractical. Omitting the parameter lets NotebookLM search the entire corpus, which naturally includes reviews + all specific papers.

When to Use Each Phase

Conversation Continuity

Use conversation_id (returned by each query) for follow-up questions. This is especially useful for the Phase 1 → Phase 2 transition:

1. Phase 1: "What are the main approaches to indirect DM detection?" → get conversation_id
2. Phase 2: "For each approach you mentioned, list specific papers with arXiv numbers" → pass conversation_id, omit source_ids

Source Provenance Classification

Every reference identified during research falls into one of two categories:

How to Determine Provenance

1. Run source_registry — this gives the full list of notebook sources with titles (usually containing arXiv IDs)
2. For each reference found during queries, check whether its arXiv ID appears as a separate notebook source filename
3. If it does → ✅ Direct Source. If not → ❌ Referenced Source

Why This Matters

When writing thesis text, we can only paraphrase and build arguments from sources we can actually read. For ❌ Referenced Sources, we rely on the review's discussion of that paper. The "Cited In" column in the data table tells us which review(s) to query for context about that paper.

External Lookup via paper_lookup Skill

For ❌ Referenced Sources, use the paper_lookup skill to retrieve additional information. Common uses:

• Quick metadata/abstract: paper_lookup Recipe 1 (InspireHEP get_paper_details)
• Full paper text: paper_lookup Recipe 2 (arXiv download_paper → read_paper)
• BibTeX entry: paper_lookup Recipe 4 (InspireHEP get_bibtex)
• Figures: paper_lookup Recipe 3 (InspireHEP get_paper_figures + download)

⚠️ Guardrails: External lookups are a triage tool, not a content source. Use them to decide whether to cite a paper, not to write based on an abstract alone. For building arguments, always rely on the review(s) that discuss the paper (listed in "Cited In"), or add the paper to the NotebookLM notebook as a full source.

Phase 3: Figure Scouting

After Phases 1–2 identify the key references, query NotebookLM for figures from the literature that could illustrate thesis content:

mcp_notebooklm_notebook_query(
    notebook_id="1b7df790-7858-4fc8-879c-39f41238c4ae",
    query="Which figures from the papers we discussed are considered
           canonical or frequently-referenced illustrations of [topic]?
           For each, state the paper (arXiv ID), figure number,
           and what it shows.",
    conversation_id=<previous_conversation_id>
)

Record figure candidates in references.md (see Figure Candidates Table in the Output section). These feed into section_drafting Step 4b and paper_lookup Recipe 3 for download.

Citation Preference Hierarchy

When citing a claim or result, follow this priority order:

1. Prefer review papers (✅ Direct Sources, 002) prefix) — they provide authoritative, synthesized context that is directly queryable
2. Complement with the specific original paper (often ❌ Referenced) — adds scholarly precision and credit to the original work
3. Use notebook general references (✅ Direct Sources without 002)) when available — these give us direct content access
4. Last resort: cite a paper only mentioned in passing in a review, without detailed discussion

Non-Peer-Reviewed Source Policy

Rule: PhD theses and unpublished preprints must NEVER be the sole citation for a specific claim. Always pair with the original peer-reviewed paper.

Example citation pattern (LaTeX):

The NFW profile~\cite{Navarro:1996gj} is the standard parametrization
for CDM halos (see~\cite{Cirelli:2024ssz} for a review).

Here Cirelli:2024ssz is ✅ (we can query it for details), while Navarro:1996gj is ❌ (cited within Cirelli). Both are cited, but the review provides processable context.

Query Best Practices

1. Granular Querying: Never ask for an entire chapter at once. Break requests by sub-section (e.g., "1.1 Cosmological Context", "1.2 Particle Nature").

2. Specific Prompt Engineering:

   • Dual-Reference Standard: Always request at least 2 distinct references for each topic — complementary perspectives (e.g., Theory vs. Observation, or two contrasting reviews).
   • Corpus Priority: Prioritize sources already in the NotebookLM corpus (review articles, books) over external textbooks from general knowledge. Only cite external sources if the notebook lacks coverage.
   • arXiv Numbers: Explicitly ask for "Specific Papers with arXiv numbers" to get primary sources for citation.
   • Relevance Justification: Ask "Why is it relevant?" to ensure each source fits the narrative.
   • Additional Sources: Always request a list of 5–8 Additional Sources per subsection, including authors, year, arXiv number, and a 5–10 word summary.

3. Follow-up Queries: Use conversation_id across both phases. For example:

   • Phase 1: "List relevant reviews on Indirect Detection" (with source_ids=<review_ids>)
   • Phase 2: "For the reviews you just listed, which specific papers do they cite for gamma-ray line searches?" (omit source_ids)

Output

1. references.md

Save to chapter_XX/references.md. Follow the structure in resources/references_structure.md:

1. Reviews & Textbooks — general consensus
2. Key Specific Papers — primary sources
3. References Breakdown by Section — detailed mapping
4. Reference Data Table — structured table with provenance tracking

2. Reference Data Table

Include as Section 4 of references.md. Format:

Column definitions:

• Bib Key: looked up from bibliography.bib (or use InspireHEP MCP). N/A = needs adding.
• In NB: ✅ = separate source in NotebookLM, ❌ = only cited within a review
• Cited In: for ❌ papers, which review(s) discuss this paper (short names: Cirelli, Hooper, Dodelson, Pinetti, Bullock, etc.)

Bib key lookup procedure:

1. Extract the arXiv number from the NotebookLM response
2. Search bibliography.bib for the arXiv number using: grep_search(query="<arxiv_number>", SearchPath="bibliography.bib", Includes=["*.bib"])
3. If found, extract the bib key (the string after @article{ or @book{ etc.)
4. If NOT found, use mcp_inspirehep_get_bibtex to fetch the entry and append to bibliography.bib
5. If InspireHEP also fails, write N/A — this signals the user needs to add the entry manually

Provenance lookup procedure:

1. Get notebook sources from source_registry output
2. For each reference, check if its arXiv ID appears in any source title
3. Mark ✅ if found, ❌ if not
4. For ❌ entries, note which review(s) cited the paper based on the query response

3. Figure Candidates Table

If Phase 3 identified relevant figures, include as Section 5 of references.md:

Column definitions:

• Figure: Figure number in the original paper
• Paper: arXiv ID (or DOI for non-arXiv papers)
• Bib Key: from bibliography.bib lookup
• Description: What the figure shows (1 line)
• Section: Which thesis section would use this figure

These candidates are consumed by section_drafting Step 4b, which attempts download via paper_lookup Recipe 3.

4. Knowledge Insights

REQUIRED: After producing references.md, use the knowledge skill (save mode) to persist key insights to .agent/knowledge/. The knowledge skill defines the standard file format (YAML frontmatter + body) and handles deduplication.

Usage Examples

Scenario: Finding General Reviews

• User: "Find reviews on Indirect Detection."
• Query: "List the most relevant review articles and books on Indirect Detection of Dark Matter. For each, explain why it is relevant."

Scenario: Finding Specific Citations

• User: "Who established the limits on neutrino masses?"
• Query: "Provide a list of specific papers establishing limits on neutrino masses (e.g., Tremaine-Gunn), including arXiv numbers and a summary of the finding."

Scenario: Defining Standard Physics

• User: "Write the equations for the NFW profile."
• Query: "Provide the mathematical definition of the NFW density profile and explain its parameters using standard references."

Scenario: Full Chapter Research

• User: "Research sources for Chapter 1."
• Action:
  1. Read outline.md to identify sub-topics for Chapter 1.
  2. Iterate through each sub-topic (1.1, 1.2, 1.3...).
  3. For each, run a query combining the strategies above.
  4. Compile into chapter_01/references.md with data table.
  5. Save key insights to .agent/knowledge/.