PaperRoach

Website · Installation guide · Issues

PaperRoach is a local-first paper knowledge pipeline for researchers who keep papers in Zotero and notes in Obsidian. It turns PDFs and Markdown notes into a linked Obsidian knowledge library, backed by LanceDB vector search and local Ollama models.

The project is designed for modest GPUs. The LLM and embedding model are never kept resident together: PaperRoach extracts and analyzes first, unloads the LLM, then loads the embedder for indexing and retrieval.

PDF papers / Markdown notes
        |
        v
PASS A0: ingest to Markdown
        |
        v
PASS A1: metadata, analysis, paper-domain classification, concepts
        |
        v
model swap: unload LLM, load embedder
        |
        v
PASS B: chunk embeddings, summary embeddings, LanceDB upsert
        |
        v
Obsidian notes, concept notes, tags, related-paper links

Features

• Ingest PDFs and Markdown notes.
• Auto-detect Zotero storage and process new attachment PDFs.
• Enrich notes from Zotero bibliographic metadata, including venue and DOI, when available.
• Generate detailed English paper notes for Obsidian.
• Classify papers into research domains and subdomains, such as Computer Science / Computer Graphics or HCI / VR/AR Interaction.
• Create merge-safe Knowledge Library concept notes.
• Maintain a controlled tag registry with aliases.
• Search and ask questions over the local library using LanceDB and Ollama.
• Keep user-written ## My Notes sections intact across rebuilds.

Requirements

• Python 3.11 or 3.12 (Python 3.12 via uv-managed Python is recommended)
• An existing Obsidian vault
• Ollama running locally at http://localhost:11434
• Zotero is optional for manual PDF builds and recommended for automatic metadata and attachment watching
• Recommended models:

ollama pull qwen3:8b
ollama pull bge-m3
# Optional, for figure_mode = "describe"
ollama pull qwen2.5vl:7b

Install

The new-user installation guide covers Windows, macOS, Linux, Obsidian and Zotero detection, Ollama model downloads, the first PDF, and troubleshooting. Install the published CLI into its own isolated environment:

uv python install 3.12
uv tool install --python 3.12 paperroach
paperroach --version

The primary CLI is paperroach. The older kb command remains as a compatibility alias. If uv reports that its tool directory is missing from PATH, run uv tool update-shell and open a new terminal.

First-time setup

paperroach setup --pull-models
paperroach doctor

Setup detects Obsidian vaults that have been opened, writes a small user-level config, prepares References/ and .kb/, reports Zotero when available, and downloads the configured Ollama models only when --pull-models is passed. It is safe to rerun. Pass --vault PATH when detection finds none or more than one. See kb/templates/kb.example.toml for every optional setting.

Configuration precedence:

CLI flags > KB_* environment variables > selected config > built-in defaults
KB_CONFIG > <explicit vault>/kb.toml > ./kb.toml > user config

If KB_CONFIG is set, it must point to an existing TOML file. Invalid TOML, boolean, and integer values fail fast with field-specific configuration errors. When --vault or KB_VAULT selects a different vault than an automatically discovered kb.toml, PaperRoach ignores that mismatched config file so another vault's absolute kb_dir or model settings are not reused by accident. Set KB_CONFIG to force a specific config file.

Generated note prose and RAG answers default to English. You can change this in kb.toml with note_language and answer_language.

Build

paperroach build "C:/papers/attention-is-all-you-need.pdf"
paperroach build "C:/papers" -r
paperroach build paper1.pdf paper2.pdf "C:/notes"

For each input, PaperRoach:

1. Extracts Markdown from the source.
2. Optionally extracts figures into vault-visible assets and describes them with a local vision model.
3. Extracts metadata and paper analysis with the LLM.
4. Classifies the paper by contribution domain.
5. Distills concepts and drafts concept-note content.
6. Chunks the document.
7. Unloads the LLM and loads the embedding model.
8. Embeds prose chunks, summaries, and figure evidence.
9. Finds related papers from the existing store and writes paper/source notes.
10. Commits successfully written documents to LanceDB and creates concept notes.
11. Saves the content-hash ledger and refreshes related-paper links.

PDFs become generated paper notes under:

<vault>/<references_dir>/<Domain>/<Subdomain>/*.md

Markdown notes are indexed for search and related-linking. Existing user notes are not overwritten; PaperRoach only manages marker-delimited blocks.

paperroach build exits with code 0 only when at least one document is successfully indexed or every remaining input is a known duplicate. If nothing is processed, automation receives a non-zero exit code.

Generated Notes

Generated paper notes use YAML frontmatter and a study-note body:

---
Date: 2026-06-22
Type:
- Paper
Status: Unread
Authors: Tianye Li et al.
Year: 2017
Domain: Computer Science
Subdomain: Computer Graphics
Secondary Domains:
- HCI
Contribution Type: method
Methods:
- statistical shape modeling
Venue: ACM Transactions on Graphics
Venue Type: journalArticle
DOI: 10.1145/example
Source: https://example.org/paper
tags: [paper, face-model]
kb-generated: true
---
# Learning a model of facial shape and expression from 4D scans

## TL;DR
...

## Problem & Motivation
...

## Approach
...

## Key Results
...

## Contributions
...

## Strengths & Limitations
...

## Concepts
- [[FLAME Model]]

## Related Papers
%% kb-related-start %%
- [[Active Shape Models-Their Training and Application (1995)]]
%% kb-related-end %%

## My Notes

Domain Classification

PaperRoach separates a paper's main contribution domain from incidental tools. For example, a VR relaxation system evaluated with participants should be filed as HCI even if it uses generative models or VR rendering. This avoids treating tool keywords as the research area.

The classifier combines:

• A controlled taxonomy in kb/taxonomy.py.
• A controlled subdomain taxonomy for nested filing and frontmatter.
• An LLM classification pass in kb/llm.py.
• A heuristic fallback when the LLM is unavailable.
• Frontmatter persistence through Domain, Secondary Domains, Subdomain, Contribution Type, and Methods.

Subdomain filing is metadata-first: explicit Subdomain frontmatter or Zotero Extra hints win, then metadata hints such as tags, venue, DOI/source URL, and title are used before the generated note body is inspected. In Zotero Extra, use lines such as PaperRoach Domain: HCI and PaperRoach Subdomain: Health & Wellbeing for explicit control.

Zotero Watcher

paperroach watch --scan
paperroach watch

The watcher auto-detects the Zotero data directory from the Zotero profile, including custom dataDir values. Override it with:

paperroach watch --zotero-dir "D:/Zotero"

New PDFs under storage/ are built once, keyed by path and content hash. Failed builds are retried on later cycles, and a shared pipeline.lock heartbeat prevents the watcher and manual write commands from racing on the same store.

When a PDF is a Zotero attachment, PaperRoach reads bibliographic fields from zotero.sqlite in read-only mode: title, authors, year, tags, URL, venue, item type, DOI, volume, issue, pages, publisher, and explicit Domain / Subdomain hints from Zotero Extra.

Query

paperroach search "self-attention computational complexity"
paperroach ask "What is the core idea behind attention in these papers?"

search uses the embedding model only. ask embeds the query, unloads the embedder, then loads the LLM to answer from retrieved context.

Maintenance

paperroach doctor
paperroach stats
paperroach relink
paperroach refile
paperroach retag
paperroach organize
paperroach gc

Useful maintenance commands:

• doctor: check config, vault/store compatibility, Zotero discovery, and Ollama reachability. Use --skip-ollama for offline checks.
• stats: show document and chunk counts.
• relink: recompute related-paper and related-concept links.
• refile: move generated paper notes into domain/subdomain folders. Add --plan-out refile-plan.md to write a reviewable Markdown move plan before applying changes.
• retag: consolidate generated note tags into the Tag Registry.
• organize: plan or apply Knowledge Library folder organization.
• gc: report or remove orphaned store rows and duplicate generated notes.

Read-only commands such as stats, search, ask, doctor, and dry-run gc return empty results or warnings without initializing .kb when the store does not exist yet.

Commands that can move or rewrite files default to dry-run mode unless --apply is provided.

Development Checks

Pull requests run the same checks on GitHub Actions for Python 3.11 and 3.12 on Linux, macOS, and Windows.

python -m unittest discover -s tests -v
python -m compileall -q kb paperroach tests
python -m pip wheel . --no-deps -w dist

The test suite includes taxonomy regression checks for metadata-first subdomain filing, pure-function edge cases, and a stubbed build -> search -> ask smoke test that exercises the real local store without requiring a live Ollama server.

Release and versioning steps are documented in RELEASE.md.

The vector store writes .kb/store_meta.json with the store schema version, embedding model, and embedding dimension. If you change embed_model or embed_dim, rebuild the store instead of reusing incompatible vectors.

Community

PaperRoach is an early public project and welcomes focused contributions. Read CONTRIBUTING.md for local setup, test expectations, and data-safety rules. Community standards are in CODE_OF_CONDUCT.md; responsible disclosure is in SECURITY.md; maintainer decision-making is described in GOVERNANCE.md.

If you use PaperRoach in research, see CITATION.cff.

PDF Parsing

The default PDF backend is pymupdf4llm on CPU. Scanned PDFs can fall back to OCR with rapidocr-onnxruntime.

Optional backends:

• docling for higher-fidelity scientific parsing.
• nougat for math-aware PDF parsing and real LaTeX equation extraction.

Set the backend in kb.toml:

ingester = "pymupdf4llm"

Or per command:

paperroach build "paper.pdf" --ingester nougat

Figure-Aware Parsing

Figure enrichment is opt-in. In extract mode PaperRoach saves useful figure crops and captions; describe additionally analyzes each crop with a local vision model. Assets live under Assets/PaperRoach/<doc-id>/, generated paper notes receive a ## Key Figures section, and visual evidence is searchable alongside prose chunks. New paper notes place up to three figure-grounded findings and inline previews directly inside the relevant Approach, Key Results, or other study-note section, so the visual evidence appears where the reader needs it.

pip install -e ".[docling]"
ollama pull qwen2.5vl:7b
paperroach build "paper.pdf" --figure-mode describe
# Fully offline embedded-image extraction:
paperroach build "paper.pdf" --figure-mode extract --figure-backend pymupdf

To enrich papers that were indexed before figure support existed, use the backfill command. It reads each generated note's original kb-source PDF, preserves the existing analysis and ## My Notes content, and updates only the generated Key Figures section and figure index. Preview first, then apply:

paperroach enrich-figures --figure-mode describe --figure-backend pymupdf
paperroach enrich-figures --figure-mode describe --figure-backend pymupdf --apply

Use --limit N to process a smaller batch. Papers with indexed figures are skipped by default; add --force to refresh them. The default command is a dry run; only --apply writes assets, notes, and LanceDB figure rows.

For notes that already have indexed figures, weave that evidence into the existing study summary without re-ingesting the PDF or rewriting the rest of the note:

paperroach integrate-figures
paperroach integrate-figures --apply

This creates or refreshes only managed inline visual-evidence blocks, preserves ## My Notes, and skips notes already integrated unless --force is supplied. The synthesis uses only indexed figure descriptions and captions; papers whose visual assets do not support a grounded claim are left unchanged.

When a PDF exposes a figure as many small image tiles rather than one large embedded crop, the PyMuPDF backend can render the compound page visual as a fallback. This path runs only when no standalone figure crop qualifies.

figure_mode = "describe"       # off | extract | describe
figure_backend = "docling"     # docling | pymupdf
vision_model = "qwen2.5vl:7b"
figure_assets_dir = "Assets/PaperRoach"
figure_max_per_paper = 12
figure_min_area_ratio = 0.02

docling is the preferred layout-aware backend: it can associate pictures, tables, captions, and page geometry. Its first use may need locally cached model artifacts. pymupdf is a fast offline fallback for PDFs that contain embedded raster images; it cannot reliably recover vector-only diagrams or complex table structure. Figure descriptions are supporting evidence for paper analysis and HCI classification, never a source for bibliographic metadata.

PaperRoach loads the vision model, then unloads it before the text LLM and embedder run, so the three models do not co-reside on an 8 GB GPU. Figure crops remain in the local vault; do not commit PDFs or extracted assets from copyrighted papers to the project repository.

Project Layout

kb/
  cli.py             command-line interface
  templates/          packaged `paperroach init` configuration template
  pipeline.py        build, watch, relink, refile, retag, gc
  ingest.py          PDF / Markdown ingestion
  figures.py         figure crops, vision evidence, and vault assets
  llm.py             LLM prompts and JSON coercion
  taxonomy.py        paper-domain taxonomy and heuristic fallback
  obsidian.py        generated notes and managed blocks
  knowledge.py       concept notes and Knowledge Library operations
  store.py           LanceDB tables and vector search
  rag.py             search and grounded question answering
paperroach/
  __main__.py        package alias for `python -m paperroach`
scripts/
  smoke_wheel.py      isolated wheel-install smoke test

License

MIT. See LICENSE.