pageindex-grounded

PageIndex Grounded Skill

Use this skill when the user needs precise, grounded follow-up answers from papers that already exist locally under the current project.

Typical use cases:

• exact claim verification from one downloaded paper
• "where does this paper say X?" questions
• page-level follow-up after fast-research, pro-research, or deep-research
• bounded cross-paper comparison across a small local paper set
• checking whether a prior synthesis overclaimed or missed a qualification

Do not use this skill for paper discovery. Use fast-research, pro-research, or deep-research first when the papers are not already local. Do not use this skill as a replacement for paper-summary. paper-summary remains the canonical per-paper summary workflow. When the user wants the figure image itself instead of only a grounded answer, hand off to the shared pdf-figure-extract skill after you identify the likely paper and page range.

Mission

Given a precise question over already-downloaded local papers:

1. Select a bounded local paper set that is likely to contain the answer.
2. Generate or reuse cached PageIndex tree artifacts for those papers.
3. Use the tree structure to retrieve the most relevant sections.
4. Verify the answer against the tree output, the local PDF, and existing paper summaries when needed.
5. Return a concise grounded answer with exact paper or page references and explicit limitations.
6. Persist reusable artifacts under the active research run folder, normally <RUN_ROOT>/pageindex/.

Prerequisites

• Local PDFs already exist under an active research run folder, normally research/<YYYY-MM-DD>-<topic-slug>/pdf/, or under the legacy flat research/pdf/ layout.
• Optional metadata exists under <RUN_ROOT>/meta/ or legacy research/meta/.
• Optional paper summaries exist under <RUN_ROOT>/pdf/*.md or legacy research/pdf/*.md.
• python3 is installed and available in PATH.
• A local checkout of the open-source PageIndex repo exists. Recommended path: tools/PageIndex.
• GEMINI_API_KEY is available for the local PageIndex runner.

If the local PageIndex checkout is missing, only install it when the user explicitly asks for installation or setup work.

Recommended local setup when the user explicitly wants installation:

git clone https://github.com/VectifyAI/PageIndex.git tools/PageIndex
python3 -m pip install -r tools/PageIndex/requirements.txt

Hard Requirements

• Operate only on already-downloaded local PDFs. Do not use this skill to search for new papers.

• Prefer selecting the active research run folder from research/INDEX.md when it exists. If there is no index, infer the best run folder from the user's topic and existing research/*/RUN.md files; fall back to legacy research/pdf/ only for older projects.

• Keep paper selection bounded before retrieval. Default to:

  • 1 paper for a single-paper question
  • 2-5 papers for a cross-paper question

• Persist PageIndex artifacts under <RUN_ROOT>/pageindex/, not in the default results/ directory.

• Maintain <RUN_ROOT>/pageindex/manifest.json so later runs can reuse existing tree artifacts.

• Prefer reusing an existing tree when the source PDF has not changed.

• The local PageIndex runner must have GEMINI_API_KEY available in the environment.

• Final answers must include exact paper or page references for non-trivial claims whenever the local evidence supports that level of grounding.

• If evidence is partial, summary-only, metadata-only, or not fully verified against the current PDF, say so explicitly.

• Default to the local open-source PageIndex workflow. Do not switch to hosted PageIndex MCP or hosted Chat API unless the user explicitly asks for that external-service path.

• OpenColab normally provides OPENCOLAB_PROGRESS_FILE during provider runs. When it is set and the task is long enough to justify updates, emit bounded JSON progress events for selection, indexing, retrieval, verification, degraded coverage, and final delivery.

OpenColab Progress Helper

OpenColab exposes this progress channel by default during provider runs. When OPENCOLAB_PROGRESS_FILE is available, use this helper:

emit_progress() {
  if [ -z "${OPENCOLAB_PROGRESS_FILE:-}" ]; then
    return 0
  fi
  printf '%s\n' "$1" >> "$OPENCOLAB_PROGRESS_FILE"
}

Write one-line JSON events. Allowed kind values are started, progress, milestone, warning, needs_input, and completed.

Example:

emit_progress '{"kind":"milestone","stage":"pageindex","slot":"grounding","message":"Selected 3 local papers for grounded retrieval."}'

Useful update categories for this skill:

• selected paper set known
• cached trees reused
• tree generation started or completed
• retrieval and verification started
• degraded run because a PDF, tree, or local PageIndex checkout is missing
• final grounded answer written

Workflow

1. Select a bounded local paper set

Use the question plus whatever local artifacts already exist:

• research/INDEX.md
• <RUN_ROOT>/RUN.md
• <RUN_ROOT>/meta/*.json
• <RUN_ROOT>/pdf/*.md
• prior <RUN_ROOT>/findings.md
• prior <RUN_ROOT>/pageindex/answers/*.md

Selection guidance:

• single-paper exact lookup: 1 paper
• "compare these two papers": 2 papers
• broader but still bounded comparison: 3-5 papers

Record the selected papers in <RUN_ROOT>/pageindex/manifest.json.

2. Prepare the PageIndex workspace

RUN_ROOT="research/<YYYY-MM-DD>-<topic-slug>"
mkdir -p "$RUN_ROOT/pageindex"/{trees,answers}

Recommended manifest shape:

{
  "generated_at": "2026-03-22T12:34:56Z",
  "papers": [
    {
      "safe_id": "arxiv__2501.01234",
      "paper_id": "arXiv:2501.01234",
      "title": "Example Paper",
      "pdf_path": "research/2026-03-22-example-topic/pdf/arxiv__2501.01234.pdf",
      "summary_path": "research/2026-03-22-example-topic/pdf/arxiv__2501.01234.md",
      "tree_path": "research/2026-03-22-example-topic/pageindex/trees/arxiv__2501.01234.json",
      "status": "indexed"
    }
  ]
}

3. Generate or refresh per-paper trees

First confirm GEMINI_API_KEY is available. If it is missing, stop and report the missing prerequisite instead of pretending the run is grounded.

For each selected paper:

python3 tools/PageIndex/run_pageindex.py \
  --pdf_path "$RUN_ROOT/pdf/<safe_id>.pdf" \
  --model gemini/gemini-3.1-flash-lite-preview \
  --if-add-node-id yes \
  --if-add-node-summary yes \
  --if-add-node-text yes

Then move or copy the generated artifact into the canonical cache path:

• from: results/<safe_id>_structure.json
• to: $RUN_ROOT/pageindex/trees/<safe_id>.json

If a cached tree already exists and the source PDF has not changed, reuse it.

4. Retrieve relevant sections with the tree

For each selected paper:

1. Read $RUN_ROOT/pageindex/trees/<safe_id>.json.
2. Use node titles, node summaries, node ids, and page ranges to shortlist relevant sections.
3. Use node text when available to narrow the answer.
4. If the question depends on exact wording, a figure, a table, or an equation, verify the relevant page or anchor against the local PDF or the existing paper-summary output. If the user wants the figure artifact returned, pass the likely page(s) to pdf-figure-extract instead of trying to answer with text alone.

For cross-paper questions, do this per paper first, then synthesize. Do not merge trees into one blob and guess.

5. Write an optional grounded answer note

When the question is non-trivial, write:

• $RUN_ROOT/pageindex/answers/<date>-<topic-slug>.md

Recommended structure:

# Grounded Answer: <topic>

## Question

...

## Selected Local Papers

- `<safe_id>` ...

## Answer

...

## Evidence

- `[Paper: <safe_id>, pp. 4-5]` ...

## Limitations

...

6. Return the final answer

The user-facing reply should:

• answer the question directly
• name the searched local paper count when it materially affects confidence
• include exact paper or page references inline or immediately after the supported claim
• surface missing PDFs, stale trees, summary-only evidence, or other limitations that materially affect confidence
• point to the saved grounded answer note when one was written

Output Contract

• <RUN_ROOT>/pageindex/manifest.json
• <RUN_ROOT>/pageindex/trees/<safe_id>.json for each indexed paper
• optional <RUN_ROOT>/pageindex/answers/<date>-<topic-slug>.md
• a concise grounded final reply with exact paper or page references when supported by the local evidence

Canonical Assets

• Skill doc: projects/SKILLS/pageindex-grounded/SKILL.md
• Expected upstream local runner: tools/PageIndex/run_pageindex.py