Paper-Reader

概述

Paper Deep Read

Three-layer academic paper analysis: Overview -> Method Detail -> Innovation.

All analysis is performed by the agent itself. The Python package handles PDF parsing and quality assessment only.

Pipeline

PDF -> parse_pdf(quality_check=True) -> Assess quality
  -> score >= 70: use extracted text
  -> score < 70 + multimodal: Read tool on PDF (VLM)
  -> score < 70 + text-only: OCR fallback
-> Layer 1 Overview -> Layer 2 Method Detail -> Layer 3 Innovation -> Markdown Report

Step 1: Parse PDF & Assess Quality

pip install pdfplumber pymupdf
python -m paper_deep_read <paper.pdf> -o parsed.json

The script always runs quality assessment (5 checks: garbled text, formula quality, text misalignment, empty sections, missing tables). Output includes quality.score (0-100).

Quality thresholds:

• >= 70: Good. Use extracted content.
• 40-69: Degraded. Switch to fallback.
• < 40: Critical. Must use fallback.

Fallback Chain

score < 70 AND model is multimodal -> Read PDF directly with Read tool (VLM understands formulas/tables natively)
score < 70 AND model is text-only  -> render pages + OCR (tencentcloud-ocr or similar)
both fail                         -> ask user

Do NOT ask user which fallback. Auto-detect model capability and proceed.

Step 2: Layer 1 – Overview

Extract in this exact order:

1. Background – Research context, motivation
2. Problem – General problem -> specific gap -> why it matters
3. Target Problem – Formal statement, input/output, constraints
4. Method – Name, category, core idea, architecture, key components
5. Experiments – Datasets, baselines, main results (with numbers)
6. Ablation – What each component contributes
7. Conclusion – Contributions, limitations, future work

Output format: Structured Markdown with tables for experiments/ablation.

Decision points:

• Language ambiguity -> match paper's primary language, ask if truly uncertain
• 10+ formulas in paper -> offer overview-first, then user-selected deep-dive

Step 3: Layer 2 – Method Detail

For every mathematical formula:

Also describe: overall architecture, data flow, training/inference pipeline, hyperparameters.

Formula template:

#### Formula [id]: [brief name]
**Formula:** [exact text]
**Purpose:** ...
**Symbols:**
| Symbol | Meaning | Type | Domain |
|--------|---------|------|--------|
| ... | ... | scalar/vector/matrix/function | R^d, ... |
**Intuition:** ...
**Connection:** links to formula [x]

Step 4: Layer 3 – Innovation & Optimization

1. Strengths (3-5) with evidence from paper
2. Weaknesses (3-5) with suggested fixes
3. Optimization Opportunities (3-5) – concrete, implementable
4. New Research Directions (3-5) – each with: title, motivation, connection, expected contribution, methodology sketch, target venues
5. Experiment Ideas – additional experiments to validate extensions

Ask user about their research direction to tailor suggestions.

Step 5: Output

1. Generate Markdown report (all 3 layers) -> save to workspace
2. Use deliver_attachments to deliver
3. Present Layer 1 inline, offer deep-dive on Layers 2/3

MCP Integration (Optional)

When available, use MCP tools for:

• Knowledge base search (related papers)
• Any tool that provides complementary analysis

Discover tools via ToolSearch with queries like ["knowledge", "search", "paper"].

Decision Points

Full decision definitions: paper_deep_read/schemas.py + paper_deep_read/prompts.py