A local-first, self-hosted research assistant inspired by NotebookLM. Built with R and Shiny.
Named after the Serapeum of Alexandria, the daughter library of the ancient Library of Alexandria.
Upload PDFs and chat with your documents using RAG (Retrieval-Augmented Generation).
- Chat with citations - Get answers with document name and page number references
- Markdown rendering - Assistant responses display with formatted headers, tables, lists, and code blocks
- One-click presets - Summarize, Key Points, Study Guide, Outline, and more
- Synthesis progress feedback - Status modal with rotating stage indicators for long-running presets
- Chat export - Download conversations as Markdown (.md) or styled HTML (.html)
- Full-text search - Vector embeddings for semantic search across documents
- Slide generation - Generate Quarto RevealJS presentations from your research
Discover and curate academic papers via OpenAlex (240M+ scholarly works).
- Smart search - Query across titles, abstracts, or full text
- Document type filters - Filter by article, review, preprint, book, dissertation, dataset
- Keyword filters - Ban or keep keywords from the global panel or directly on any abstract card; promoted keywords appear in the global chip bin
- Quality filters - Exclude retracted papers, flag predatory journals/publishers
- Citation filters - Set minimum citation thresholds
- Rich metadata display:
- Type badges (article, review, preprint, etc.)
- Open Access status badges (gold, green, hybrid, bronze, closed)
- Citation metrics (cited-by count, FWCI, reference count)
- Paper keywords from OpenAlex
- DOI as clickable link (with citation key fallback for legacy papers)
- Citation export - Download results as BibTeX (.bib) or CSV (.csv) with unique citation keys
- Chat export - Download abstract chat conversations as Markdown or HTML
- Export to seed search - Use any paper as a seed for a new discovery search with one click
- Import to documents - Move curated papers to document notebooks for deeper analysis
Extract, describe, and manage figures from academic PDFs.
- Automatic extraction - Renders PDF pages, detects text gaps, and crops figure regions
- Caption detection - Associates extracted captions with their figures
- AI descriptions - Vision model (GPT-4.1 Nano) generates structured descriptions for each figure
- Figure gallery - Review figures in list view (detailed) or thumbnail grid (quick scan)
- Per-figure actions - Keep, Ban (exclude from slides), or Retry (re-describe) each figure
- Watermark filtering - Automatically removes publisher watermarks that interfere with extraction
- Cost tracked - Vision API costs logged alongside other LLM usage
Generate presentation slides from notebook content using Quarto RevealJS.
- Configurable options - Length, audience level, theme selection
- 11 RevealJS themes - moon, sky, beige, serif, and more
- Figure integration - Include extracted figures in slides with layout-aware placement
- Aspect-ratio guidance - Wide, standard, and tall figures get appropriate slide layouts
- Speaker notes - Optional auto-generated presenter notes
- Multiple formats - Preview in-app, download .qmd, export to HTML/PDF
- Custom instructions - Guide the AI on focus areas
Explore citation relationships through interactive network graphs.
- Multi-seed networks - Seed from all papers in a notebook or BibTeX import
- Overlap detection - Papers cited by multiple seeds highlighted as diamonds
- Shape encoding - Stars (seeds), diamonds (overlap), dots (regular) with year color gradient
- Missing papers discovery - Find frequently-cited papers not in your collection, import with one click
- Directional control - Explore forward citations, backward references, or both
- Configurable depth - Traverse 1-3 hops from seed papers
- Node cap - Per-seed node limits to keep graphs readable
- Interactive graph - Pan, zoom, click nodes to view paper details
- Physics toggle - Freeze/unfreeze node simulation from the legend panel
- Color palettes - Five viridis color schemes with live-switching
- Save & reload - Persist networks to database with layout positions preserved
- Collapsible legend - Minimizable legend with shape key and gradient preview
Score, rank, and curate papers against your research focus.
- Multiple anchor types - Start from seed papers, a research intent, an entire notebook, or a combination
- Tier 1 scoring - Citation velocity, FWCI, seed connectivity, bridge score, ubiquity penalty
- Tier 2 semantic scoring - BM25 + vector similarity via ragnar against embedded notebooks
- Scoring modes - Discovery (novel connections), Comprehensive (balanced), Emerging (recent high-velocity)
- Advanced weights - Fine-tune component weights with sliders; auto-normalized to sum to 1.0
- Curation workflow - Accept/reject individual papers, batch accept top N, reject bottom half
- Import results - Send curated papers to an existing or new notebook
- Notebook anchor - Use all papers in a notebook as seeds, fetch related candidates from OpenAlex with configurable per-seed limits
Full dark mode support with Catppuccin color palette.
- One-click toggle - Moon/sun icon in the navbar switches between light (Latte) and dark (Mocha) themes
- Persistent preference - Theme choice saved to localStorage, restored on reload
- Auto-themed plots - Chart backgrounds adapt automatically via thematic integration
- Comprehensive coverage - All components styled: value boxes, alerts, chat messages, network graphs, tables
Monitor API usage and performance in real-time.
- Session costs - Live cost display in the sidebar footer
- OpenRouter balance - View remaining credits and usage
- Cost history - 30-day bar chart of daily API spending with operation breakdown
- Per-call breakdown - Detailed log of every API call with model, tokens, and cost
- Latency analytics - Average, p50, p95 latency per model and operation type
- Daily latency trend - 30-day sparkline of response times
Flexible model routing with multi-provider support.
- 3-slot model routing - Assign different models to fast (query building), quality (chat, synthesis), and embedding operations
- Multi-provider support - Use OpenRouter alongside local providers (Ollama, LM Studio, vLLM)
- Provider management - Add, edit, test, and delete OpenAI-compatible endpoints from Settings
- Model benchmarks - Quality scores, speed, and pricing from Artificial Analytics
- Smart defaults - Suggests optimal models for each slot based on benchmark data
- Zero-cost local models - Local provider calls tracked at $0 with graceful handling of missing usage tokens
- Embedding dimension detection - Warns when switching to a model with different dimensions
- Guided welcome wizard - 5-step workflow modal for new users (Set Up → Find → Collect → Analyze → Audit)
- Welcome landing page - Persistent home screen with live setup status indicators and action buttons
- Contextual help text - Each sidebar section shows a brief description on first load
- Version tracking - Version badge in the title bar, What's New section on the About page
- API key validation - Visual indicators show if keys are configured and working
- Model selection - Choose quality, fast, and embedding models with AA benchmark enrichment
- Providers - Manage multiple OpenAI-compatible API endpoints
- Quality data downloads - Fetch predatory journal lists and retraction databases
- Verbose API logging - Toggle to log OpenAlex API calls to the R console for debugging
- All data stays local - DuckDB for portable, single-file storage
- No cloud dependencies - Everything runs on your machine
- Single-user - No authentication needed
- Portable - Copy the database file to move your research
- R (>= 4.6)
- Quarto (for slide generation)
- RStudio (optional but recommended)
# Clone the repository
git clone https://github.com/seanthimons/serapeum.git
cd serapeum
# One-shot setup: restores all R packages with pak and Posit Package Manager
Rscript setup.RThe committed boosterpak setup configures binary installs through Posit Package Manager and pak automatically. No manual package management is needed.
Configure API keys via the Settings page in the app (recommended), or copy the example config:
cp config.example.yml config.ymlopenrouter:
api_key: "your-openrouter-key" # Get from openrouter.ai/keys
openalex:
email: "your@email.com" # For polite pool access (faster rate limits)Serapeum can use local models via any OpenAI-compatible server instead of (or alongside) OpenRouter. This means your data never leaves your machine.
Supported servers: Ollama, LM Studio, vLLM, or any OpenAI-compatible endpoint.
- Install LM Studio and download a chat model (e.g., Gemma 3, Llama 3, Qwen 2.5)
- Optionally download an embedding model (e.g.,
nomic-embed-text-v1.5) - Start the local server (Developer tab → Start Server, default port 1234)
- In Serapeum Settings → Providers, click Add Provider:
- Name:
LM Studio - Base URL:
http://localhost:1234/v1 - API Key: leave blank
- Name:
- Click Test to verify connectivity
- In Settings → Models, select your local models for each slot:
- Quality model: your chat model
- Fast model: same chat model (or a smaller one)
- Embedding model: your embedding model
- Install Ollama and pull models:
ollama pull llama3.2 ollama pull nomic-embed-text
- In Serapeum Settings → Providers, click Add Provider:
- Name:
Ollama - Base URL:
http://localhost:11434/v1 - API Key: leave blank
- Name:
- Select your models in Settings → Models
Local models are tracked at $0.00 in the Cost Tracker. No OpenRouter API key is needed if you only use local models.
shiny::runApp()Open http://localhost:8080 in your browser.
- DuckDB database is created automatically
- Quality data (predatory journals, retraction watch, OpenAlex topics) is seeded from bundled RDS files — no download needed
- Welcome wizard guides you through setup, search, and analysis in 5 steps
- Click "New Document Notebook"
- Give it a name
- Upload PDFs using the upload button
- Wait for processing (text extraction)
- Click "Embed Documents" to generate embeddings
- Ask questions in the chat interface
- Use preset buttons for common tasks (Summary, Key Points, etc.)
- Click the image icon on a PDF to extract figures — review in gallery, ban unwanted ones
- Generate slides with the "Slides" tab
- Click "New Search Notebook"
- Enter a search query and configure filters:
- Date range
- Document types (article, review, preprint, etc.)
- Open access only
- Minimum citations
- Exclude retracted papers
- Click "Refresh" to search OpenAlex
- Browse results - each paper shows:
- Type badge (article, review, etc.)
- OA status badge (gold, green, hybrid, etc.)
- Citation metrics (cited-by, FWCI, references)
- Keywords
- Remove unwanted papers with the X button
- Click "Embed Papers" to enable semantic search
- Query the abstracts in chat
- Export results: Export dropdown → BibTeX (.bib) or CSV (.csv)
- Use "Use as Seed" on any paper to launch a new discovery search
- Import selected papers to a document notebook
- API Keys - Configure OpenRouter and OpenAlex credentials
- Visual indicators show validation status (green check = valid)
- Models - Select chat and embedding models
- Quality Data - Download predatory journal/publisher lists and retraction database
- DOI Management - View DOI coverage stats and backfill missing DOIs from OpenAlex
- R + Shiny + bslib: Web framework with Bootstrap 5 UI components
- DuckDB: Embedded analytical database for local storage
- OpenRouter: Unified API for multiple LLM providers (Claude, GPT-4, Llama, etc.)
- Provider Abstraction: OpenAI-compatible endpoint support for local models (Ollama, LM Studio, vLLM)
- Artificial Analytics: Model benchmark data for quality/speed/price comparison
- OpenAlex: Free, open academic paper search API
- Quarto: Scientific publishing system for slide generation
- pdftools: PDF text extraction and page rendering (bundles Poppler)
- base64enc: Image encoding for vision API
- digest: Figure deduplication
serapeum/
├── app.R # Main Shiny app
├── setup.R # One-shot setup script
├── .Rprofile # Auto-activates renv
├── renv.lock # Locked dependency versions
├── config.yml # Your config (gitignored)
├── config.example.yml # Config template
├── R/
│ ├── config.R # Config loading
│ ├── db.R # Database operations
│ ├── db_migrations.R # Schema migrations
│ ├── api_openrouter.R # OpenRouter client
│ ├── api_openalex.R # OpenAlex client
│ ├── api_provider.R # Provider abstraction layer
│ ├── api_artificialanalysis.R # AA benchmarks client
│ ├── pdf.R # PDF text extraction
│ ├── pdf_extraction.R # Figure extraction from PDFs
│ ├── pdf_images.R # Figure storage, vision description, orchestrator
│ ├── rag.R # RAG pipeline
│ ├── _ragnar.R # Ragnar embedding store helpers
│ ├── slides.R # Slide generation
│ ├── cost_tracking.R # API cost tracking and pricing
│ ├── theme_catppuccin.R # Catppuccin Latte/Mocha dark mode CSS
│ ├── citation_network.R # Citation graph data and layout
│ ├── quality_filter.R # Predatory/retraction filtering + auto-seed
│ ├── interrupt.R # Graceful request cancellation
│ ├── utils_doi.R # DOI normalization and citation keys
│ ├── utils_citation.R # BibTeX/CSV export formatters
│ ├── utils_export.R # Chat export formatters (Markdown/HTML)
│ ├── utils_filters.R # Search filter utilities
│ ├── mod_about.R # About page
│ ├── mod_citation_network.R # Network visualization UI
│ ├── mod_cost_tracker.R # Cost tracking dashboard
│ ├── mod_document_notebook.R
│ ├── mod_search_notebook.R
│ ├── mod_seed_discovery.R # Seed paper discovery
│ ├── mod_query_builder.R # LLM-assisted query building
│ ├── mod_topic_explorer.R # Topic browsing
│ ├── mod_journal_filter.R # Journal filtering
│ ├── mod_keyword_filter.R # Keyword filtering
│ ├── mod_bulk_import.R # DOI/BibTeX bulk import
│ ├── mod_settings.R
│ └── mod_slides.R
├── data/
│ ├── support/ # Bundled RDS files (quality data, topics)
│ ├── figures/ # Extracted figure PNGs (auto-created)
│ └── notebooks.duckdb # Database file (auto-created)
├── storage/ # Uploaded PDFs
├── output/ # Generated slides
└── tests/
└── testthat/ # Unit tests
testthat::test_dir("tests/testthat")Delete data/notebooks.duckdb to start fresh.
We welcome contributions! Please see:
- CONTRIBUTING.md for detailed contribution guidelines
- CODE_OF_CONDUCT.md for community standards
- TODO.md for the feature roadmap and open issues
Important: Serapeum is a research tool powered by AI language models.
- Not an Oracle: AI-generated responses may contain errors, hallucinations, or inaccuracies. Always verify important information from primary sources.
- Not Professional Advice: This tool is not a substitute for professional, medical, legal, financial, or other expert advice.
- Makes Mistakes: AI models can misinterpret documents, generate plausible-sounding but incorrect answers, and miss important context.
- Research Tool Only: Intended for exploratory research and learning. Critical decisions should be based on careful review of original sources.
MIT - Copyright (c) 2024-2026 Sean Thimons
To report security vulnerabilities, please see SECURITY.md.
- Inspired by NotebookLM
- Paper data from OpenAlex
- LLM access via OpenRouter
- Quality data from Retraction Watch and Predatory Journals