research-pipeline 0.12.6

A multi-source pipeline for searching, screening, downloading, converting, and summarizing academic papers from arXiv, Google Scholar, and more.

research-pipeline

A production-grade, deterministic Python pipeline for searching, screening, downloading, converting, and summarizing academic papers from arXiv, Google Scholar, Semantic Scholar, OpenAlex, and DBLP.

Features

• Multi-stage pipeline: plan → search → screen → download → convert → extract → summarize
• 5 new auxiliary commands: expand (citation graph), quality (evaluation scoring), convert-rough / convert-fine (two-tier conversion), index (incremental runs)
• Modular CLI with independent, composable stage commands
• MCP server for AI agent integration (17 tools, 15 resources, 6 prompts, completions, progress reporting)
• Harness-engineered research workflow — server-driven orchestration with 6 layers: telemetry, context engineering, governance, structural verification, doom-loop monitoring, and crash recovery
• Multi-source search: arXiv + Google Scholar + Semantic Scholar + OpenAlex + DBLP
• Cross-source enrichment — fill missing abstracts via DOI lookup
• Semantic re-ranking — optional SPECTER2 embeddings for similarity scoring
• Citation graph expansion — discover related papers via Semantic Scholar citations
• Quality evaluation — composite scoring: citation impact, venue reputation, author h-index, recency
• Multi-backend PDF conversion: 3 local (Docling, Marker, PyMuPDF4LLM) + 5 cloud (Mathpix, Datalab, LlamaParse, Mistral OCR, OpenAI Vision)
• Two-tier conversion — fast convert-rough for all papers, high-quality convert-fine for selected ones
• Multi-account rotation — rotate between accounts per service on quota exhaustion
• Cross-service fallback — automatic failover to next backend when all accounts are exhausted
• Incremental runs — SQLite global index deduplicates papers across runs
• Retry & error recovery — @retry decorator with exponential backoff, jitter, and Retry-After support
• Idempotent & resumable — every stage can be re-run safely
• arXiv polite-mode — strict rate limiting, single connection, caching
• Deterministic tool chain with optional LLM judgment
• Full artifact lineage — every run is reproducible and auditable via manifests
• Offline-first testing — no live API calls in CI

Installation

# From PyPI
pip install research-pipeline

# With local PDF conversion backends
pip install research-pipeline[docling]       # MIT license, great tables/equations
pip install research-pipeline[marker]        # Highest accuracy (95.7%), GPL-3.0
pip install research-pipeline[pymupdf4llm]   # Fastest (10-50x), AGPL

# With cloud PDF conversion backends (require API keys)
pip install research-pipeline[mathpix]       # Best LaTeX, 1K free pages/mo
pip install research-pipeline[datalab]       # Hosted Marker, $5 free credit
pip install research-pipeline[llamaparse]    # 1K free pages/day
pip install research-pipeline[mistral-ocr]   # Mistral OCR, free credits
pip install research-pipeline[openai-vision] # GPT-4o vision

# With Google Scholar support
pip install research-pipeline[scholar]

# With all extras
pip install research-pipeline[docling,marker,pymupdf4llm,scholar]

Development install

# With uv (recommended)
uv sync --extra dev --extra docling --extra scholar

Quick start

# Full end-to-end pipeline
research-pipeline run "transformer architectures for time series forecasting"

# Or run stages individually
research-pipeline plan "transformer architectures for time series forecasting"
research-pipeline search --run-id <RUN_ID>
research-pipeline screen --run-id <RUN_ID>
research-pipeline download --run-id <RUN_ID>
research-pipeline convert --run-id <RUN_ID>
research-pipeline extract --run-id <RUN_ID>
research-pipeline summarize --run-id <RUN_ID>

# Inspect run status
research-pipeline inspect --run-id <RUN_ID>

# Standalone PDF conversion (no workspace required)
research-pipeline convert-file paper.pdf -o paper.md

# Use a specific conversion backend
research-pipeline convert --run-id <RUN_ID> --backend marker
research-pipeline convert-file paper.pdf --backend pymupdf4llm

# Two-tier conversion: rough (fast) then fine (high-quality)
research-pipeline convert-rough --run-id <RUN_ID>
research-pipeline convert-fine --run-id <RUN_ID>

# Evaluate paper quality (citation impact, venue, author)
research-pipeline quality --run-id <RUN_ID>

# Expand via citation graph (Semantic Scholar)
research-pipeline expand --run-id <RUN_ID> --direction both

# Manage global paper index (incremental dedup)
research-pipeline index --list

Commands

Command	Purpose
plan	Normalize topic → structured query plan
search	Execute multi-source search (arXiv, Scholar, Semantic Scholar, OpenAlex, DBLP)
screen	Two-stage relevance filtering (BM25 + optional SPECTER2 + optional LLM)
download	Download shortlisted PDFs with rate limiting and retry
convert	PDF → Markdown (8 backends, multi-account rotation, cross-service fallback)
convert-rough	Fast Tier 2 conversion (pymupdf4llm) for all downloaded PDFs
convert-fine	High-quality Tier 3 conversion for selected papers
extract	Structured content extraction & chunking
summarize	Per-paper summaries + cross-paper synthesis
expand	Citation graph expansion via Semantic Scholar API
quality	Composite quality evaluation (citations, venue, author, recency)
run	End-to-end orchestration of all stages
inspect	View run manifests and artifacts
convert-file	Standalone PDF → Markdown conversion
index	Manage the global paper index for incremental runs
setup	Install skill + sub-agent definitions to ~/.claude/

MCP server

The MCP server provides full Model Context Protocol support for AI agent integration:

# Run via module
uv run python -m mcp_server

17 tools — all pipeline stages plus auxiliary commands and workflow:

plan_topic, search, screen_candidates, download_pdfs, convert_pdfs, extract_content, summarize_papers, run_pipeline, get_run_manifest, convert_file, list_backends, expand_citations, evaluate_quality, convert_rough, convert_fine, manage_index, research_workflow

15 resources — read pipeline artifacts via URI templates:

runs://list, runs://{run_id}/manifest, runs://{run_id}/plan, runs://{run_id}/candidates, runs://{run_id}/shortlist, runs://{run_id}/papers/{paper_id}, runs://{run_id}/markdown/{paper_id}, runs://{run_id}/summary/{paper_id}, runs://{run_id}/synthesis, runs://{run_id}/quality, config://current, index://papers, workflow://{run_id}/state, workflow://{run_id}/telemetry, workflow://{run_id}/budget

6 prompts — research workflow templates:

research_topic, research_workflow, analyze_paper, compare_papers, refine_search, quality_assessment

Plus: tool annotations, auto-completions, and progress reporting.

Harness-engineered workflow

The research_workflow tool drives a server-side orchestrated research workflow with 6 harness engineering layers derived from a 79-paper synthesis:

Layer	Purpose
WL1 Telemetry	Three-surface logging (cognitive/operational/contextual)
WL2 Context	Token budgets, 5-stage paper compaction (Tokalator/ACC)
WL3 Governance	Schema-level state machine, verify-before-commit gates
WL4 Verification	Structural output validation (not self-referential)
WL5 Monitoring	Doom-loop detection, iteration drift tracking
WL6 Recovery	Persistent state after every stage, crash-recovery

Features:

• Sampling-based analysis: LLM paper analysis via create_message() (1 round per paper)
• Elicitation gates: user approval at 6 decision points via ctx.elicit()
• Iterative synthesis: system-building mode with gap analysis and convergence
• Bounded rationality: max 3 iterations, 7 explicit stop conditions
• Graceful degradation: works without sampling or elicitation capabilities

AI skill & sub-agents

Install the bundled Claude Code / GitHub Copilot skill and sub-agent definitions:

# Install everything (skill + agents) to ~/.claude/
research-pipeline setup

# Or create symlinks (for development)
research-pipeline setup --symlink

# Force overwrite existing
research-pipeline setup --force

# Install only the skill (skip agents)
research-pipeline setup --skip-agents

# Install only agents (skip skill)
research-pipeline setup --skip-skill

This installs:

• Skill → ~/.claude/skills/research-pipeline/ (SKILL.md + references + config)
• Sub-agents → ~/.claude/agents/ (paper-analyzer, paper-screener, paper-synthesizer)

Configuration

Copy config.example.toml to config.toml and adjust settings:

cp config.example.toml config.toml

Key environment variables:

Variable	Purpose
ARXIV_PAPER_PIPELINE_CONFIG	Config file path
ARXIV_PAPER_PIPELINE_CACHE_DIR	Override cache directory
ARXIV_PAPER_PIPELINE_WORKSPACE	Override workspace directory
ARXIV_PAPER_PIPELINE_DISABLE_LLM	Force LLM off

Artifact layout

Each pipeline run produces outputs in runs/<run_id>/:

runs/<run_id>/
├── run_config.json            # Configuration snapshot
├── run_manifest.json          # Execution metadata & stage records
├── plan/query_plan.json       # Normalized query plan
├── search/
│   ├── raw/*.xml              # Raw API response pages
│   └── candidates.jsonl       # Deduplicated candidates
├── screen/
│   ├── cheap_scores.jsonl     # Heuristic scores
│   └── shortlist.json         # Papers selected for download
├── download/
│   ├── pdf/*.pdf              # Downloaded papers
│   └── download_manifest.jsonl
├── convert/
│   ├── markdown/*.md          # Converted Markdown
│   └── convert_manifest.jsonl
├── convert_rough/             # Tier 2: fast conversion (all PDFs)
│   ├── markdown/*.md
│   └── convert_manifest.jsonl
├── convert_fine/              # Tier 3: high-quality conversion (selected)
│   ├── markdown/*.md
│   └── convert_manifest.jsonl
├── quality/                   # Quality evaluation scores
│   └── quality_scores.jsonl
├── expand/                    # Citation graph expansion
│   └── expanded_candidates.jsonl
├── extract/*.extract.json     # Chunked & indexed extraction
├── summarize/
│   ├── *.summary.json         # Per-paper summaries
│   ├── synthesis.json         # Cross-paper synthesis
│   └── synthesis.md           # Human-readable synthesis
├── workflow/                  # Harness-engineered workflow state
│   ├── state.json             # Workflow state (stage statuses, execution log)
│   └── telemetry.jsonl        # Three-surface telemetry events
└── logs/pipeline.jsonl        # Structured logs

Development

# Install dev dependencies
uv sync --extra dev

# Run unit tests
uv run pytest tests/unit/ -xvs

# Format, lint, type check
uv run isort . && uv run black . && uv run ruff check . --fix
uv run mypy src/

# Run all pre-commit hooks
uv run pre-commit run --all-files

See docs/architecture.md for detailed architecture documentation and docs/user-guide.md for the full user guide.

License

MIT