paperhound 0.5.1

CLI to search, download, and convert academic papers (arXiv, OpenAlex, DBLP, Crossref, Hugging Face Papers, Semantic Scholar, CORE) into Markdown — built for AI/ML researchers.

paperhound

paperhound — sniff out academic papers from Python or the command line.

A small, fast Python library (and matching CLI) for AI/ML researchers and tooling authors who want a single dependency to search, inspect, download, and convert to Markdown papers from many academic sources at once. Conversion is powered by docling, so the resulting Markdown is good enough to feed straight into an LLM context window.

paperhound is built primarily as a library — every CLI command is a thin wrapper around a public Python API you can import directly. The CLI is the fastest way to drive it from the terminal or from agents, but anything you can do at the prompt you can also do with three lines of Python.

Features

• 🔎 Unified search — one query, many backends. arXiv, OpenAlex, DBLP, Crossref and Hugging Face Papers (and optionally Semantic Scholar / CORE) are queried in parallel with a 10-second budget. Results are merged round-robin (one from each provider, then the next, …) so a fast provider can't monopolize the top-N — and deduplicated by arXiv id / DOI / title. Slow providers are dropped silently — you get whatever came back in time.
• 📄 Inspect before downloading — fetch the abstract and metadata, decide if the paper is worth the bytes.
• ⬇️ Download by identifier — arXiv id, DOI, Semantic Scholar paper id, or any paper URL. Open-access PDFs are resolved automatically.
• 📝 PDF → Markdown via docling — figures, LaTeX equations and HTML tables are all opt-in flags.
• 📚 Local library — a SQLite FTS5 database at ~/.paperhound/library/. Add a paper, store its Markdown, then offline-grep over titles, abstracts and bodies.
• 🧠 Optional embedding rerank — install paperhound[rerank] and the CLI reranks results by query/abstract similarity automatically.
• 🤖 Agent-ready CLI — every command speaks JSON via --json, and the repo ships a skills.sh skill so any Claude / OpenAI / local agent can drive paperhound with no extra glue.
• 🧪 Heavily tested — every module has unit tests; live integration tests sit under tests/integration/.

Installation

pip install paperhound

or with uv:

# As a library inside another project
uv add paperhound

# Or as an isolated CLI on your $PATH
uv tool install paperhound

Python 3.10+ is required. Docling pulls in PyTorch on first run, so the very first conversion may take a moment to download model weights.

Optional: embedding rerank

pip install 'paperhound[rerank]'

Adds sentence-transformers so paperhound can rerank merged search results by embedding similarity between the query and each candidate's title + abstract. When the extra is installed, rerank runs automatically on every CLI search (--no-rerank to skip). Library callers opt in by calling paperhound.rerank.rerank(...) themselves.

Quick start (Python)

from paperhound import search_papers, get_paper, convert_to_markdown

# 1. Search across all default providers (arxiv + openalex + dblp + crossref + hf).
papers = search_papers("retrieval augmented generation", limit=5)
for p in papers:
    print(p.year, p.title, p.identifiers.arxiv_id or p.identifiers.doi)

# 2. Pull the abstract for a single paper.
paper = get_paper("1706.03762")
print(paper.title)
print(paper.abstract)

# 3. PDF → Markdown (path or URL).
md = convert_to_markdown("https://arxiv.org/pdf/1706.03762", output="attention.md")

Every function returns a typed paperhound.models.Paper object — pydantic under the hood, so you get autocomplete, model_dump(mode="json"), validation, and stable field names across providers.

Building a corpus

from paperhound import Library, search_papers, get_paper

lib = Library()                   # ~/.paperhound/library/ by default
for hit in search_papers("vision language models", limit=20):
    lib.add(hit)                  # idempotent — re-adds update in place

# Offline full-text search later, no API calls.
hits = lib.grep("multi-head attention", limit=10)
for h in hits:
    print(h.id, h.title)

Citation graph

from paperhound.citations import fetch_references, fetch_citations

refs   = fetch_references("1706.03762", depth=1, limit=10)
citing = fetch_citations("1706.03762", depth=2, limit=50)

Library API reference

Top-level package re-exports the symbols you need most often:

Symbol	Purpose
paperhound.search_papers(query, limit=10, sources=None, **filters)	Run a unified search; returns list[Paper].
paperhound.get_paper(identifier)	Resolve an id/DOI/URL to a single Paper, or None.
paperhound.convert_to_markdown(src, output=None, options=None)	PDF/URL → Markdown via docling.
paperhound.pdf_to_markdown(...)	Lower-level PDF-only entry point.
paperhound.Paper, Author, PaperIdentifier	Pydantic models.
paperhound.Library	SQLite FTS5 library wrapper.

Need finer control? Drop into the underlying modules:

• paperhound.search — provider registry, SearchAggregator, SearchQuery, SearchProvider base class. Plug in your own provider with one register("name", Factory) call.
• paperhound.download — resolve_pdf_url, download_pdf.
• paperhound.convert — ConversionOptions, convert_to_markdown.
• paperhound.citations — fetch_references, fetch_citations.
• paperhound.rerank — optional, requires the rerank extra.
• paperhound.errors — PaperhoundError, ProviderError, LibraryError, RerankError. Every other exception bubbles untouched.

CLI

Once installed, paperhound is on your $PATH.

# Search across all providers
paperhound search "diffusion transformers" --limit 5

# Show the abstract for a specific paper
paperhound show 2401.12345
paperhound show 10.1038/s41586-020-2649-2          # DOI works too
paperhound show https://arxiv.org/abs/1706.03762   # ...and URLs
paperhound show 2001.08361 -s arxiv                # force a single provider

# Download the PDF
paperhound download 1706.03762 -o ./papers/

# Convert a local PDF to Markdown
paperhound convert ./papers/1706.03762.pdf -o attention.md

# Or do it all at once: search-resolve, download, convert, clean up
paperhound get 1706.03762 -o attention.md

JSON output for scripts and agents

--json is the pipe-friendly mode: no headers, no Rich formatting, no progress bars.

# search --json: JSONL — one compact JSON object per line (Paper schema)
paperhound search "graph neural networks" --json | jq '.title'

# show --json: single compact JSON object on one line
paperhound show 1706.03762 --json | jq .abstract

The schema is paperhound.models.Paper serialised via model_dump(mode="json"). Fields: title, authors[], abstract, year, venue, url, pdf_url, citation_count, identifiers.{arxiv_id,doi,semantic_scholar_id,openalex_id, dblp_key,core_id}, sources[].

--json and --format are mutually exclusive on show — use one or the other.

Commands

Command	Description
paperhound search <query>	Run a unified search. --limit, --source arxiv|openalex|dblp|crossref|huggingface|semantic_scholar|core (repeatable), --year RANGE, --min-citations N, --venue STRING, --author STRING, --timeout, --json (JSONL output), --rerank/--no-rerank (default on when paperhound[rerank] is installed), --rerank-model.
paperhound show <id>	Fetch a paper's metadata + abstract. --source arxiv|openalex|dblp|crossref|huggingface|semantic_scholar|core (-s, repeatable — restrict the lookup, useful when an upstream aggregator returns poisoned metadata), --format markdown|bibtex|ris|csljson (default markdown), --json (compact JSON; mutually exclusive with --format).
paperhound download <id> -o <path>	Download a paper PDF.
paperhound convert <pdf> -o <md>	Convert a PDF (or any docling-supported file/URL) to Markdown. --with-figures saves embedded images to <stem>_assets/ and references them in the output. --equations latex preserves math as $...$/$$...$$. --tables html embeds <table> blocks instead of GFM pipe tables.
paperhound get <id> -o <md>	Download + convert in one step. --keep-pdf to keep the PDF.
paperhound refs <id>	List works the paper cites (its references). --depth 1|2, --limit N, --source openalex|semantic_scholar, --json.
paperhound cited-by <id>	List works that cite the paper. Same flags as refs.
paperhound add <id>	Fetch metadata and add to local library. --convert also stores Markdown.
paperhound list	List all papers in the local library.
paperhound grep <query>	Full-text search the local library (title + abstract + Markdown body).
paperhound rm <id>	Remove a paper from the local library (and its Markdown file, if any).
paperhound version	Print the installed version.

Run paperhound <command> --help for full options.

Conversion options

paperhound convert (and the get / add --convert pipeline) accepts three flags that control how the PDF is rendered to Markdown:

Flag	Values	Default	Description
--with-figures	—	off	Extract embedded figures to <stem>_assets/ and embed  references. Requires --output.
--equations	inline, latex	inline	latex enables formula enrichment — math is preserved as $...$ / $$...$$ LaTeX (uses docling's do_formula_enrichment VLM; slightly slower).
--tables	markdown, html	markdown	html embeds raw <table> blocks for better fidelity with merged/irregular cells.

paperhound convert paper.pdf -o paper.md --with-figures --equations latex --tables html
paperhound convert paper.pdf -o paper.md --equations latex
paperhound convert paper.pdf -o paper.md --tables html

All three flags default to the original behaviour, so existing pipelines are unaffected.

Filters

paperhound search accepts four filter flags. Filters are pushed down to providers that support them (OpenAlex, Crossref, Semantic Scholar) and always applied client-side after the merge as a safety net.

Flag	Accepted values	Example
--year RANGE	YYYY, YYYY-YYYY, YYYY-, -YYYY	--year 2022-2024
--min-citations N	integer ≥ 0	--min-citations 100
--venue STRING	case-insensitive substring	--venue NeurIPS
--author STRING	case-insensitive substring	--author Hinton

paperhound search "vision transformers" --year 2022-2024 --min-citations 100
paperhound search "deep learning" --venue NeurIPS --author Hinton
paperhound search "diffusion models" -s arxiv --year 2023-
paperhound search "llm alignment" --year 2023 --min-citations 50 --json | jq .title

Behavior with missing fields: papers whose year or venue field is unknown (null) are kept — the filter cannot be verified. Papers whose citation_count is unknown are excluded when --min-citations is set (conservative: the user asked for a floor).

Export formats

paperhound show can export a paper's metadata in four formats:

paperhound show 1706.03762                       # rich terminal view
paperhound show 1706.03762 --format bibtex
paperhound show 1706.03762 --format ris
paperhound show 1706.03762 --format csljson

BibTeX cite keys are derived deterministically as <firstAuthorLastName><year><firstSignificantTitleWord> (accents stripped, lowercased). LaTeX special characters (&, %, $, _, etc.) are escaped automatically.

Local library

paperhound keeps a persistent per-user library at ~/.paperhound/library/ (override with PAPERHOUND_LIBRARY_DIR). The library is backed by a SQLite FTS5 database — no extra dependencies required.

paperhound add 1706.03762
paperhound add 1706.03762 --convert
paperhound list
paperhound grep "attention mechanism"
paperhound rm 1706.03762

Re-adding a paper is idempotent — it updates the metadata in place. The schema is versioned; on a version mismatch paperhound reports a clear error rather than silently operating on a stale schema.

Citation graph

paperhound refs 1706.03762
paperhound cited-by 1706.03762
paperhound refs 1706.03762 --depth 2 --limit 50
paperhound cited-by 1706.03762 --source semantic_scholar
paperhound refs 1706.03762 --json | jq '.[].title'

Both commands return the same Paper format as search. The default provider order is OpenAlex first, Semantic Scholar as fallback (automatically triggered when OpenAlex returns nothing or errors). Results are deduplicated by arXiv id / DOI / title before being returned. At --depth 2, total fetched is capped at limit * 2 and a small pause (0.1 s) is inserted between hops to stay in the polite API pool.

Rerank

When paperhound[rerank] is installed, every CLI search call reranks results by embedding similarity between the query and each candidate's title + abstract. Pass --no-rerank to skip it for one call.

pip install 'paperhound[rerank]'

paperhound search "vision language models"          # rerank on by default
paperhound search "graph neural networks" --no-rerank
paperhound search "agents" --rerank-model sentence-transformers/all-mpnet-base-v2

Without the extra installed the CLI silently falls back to the merge-order ranking — no error, no hang. Library users that want to invoke rerank directly call paperhound.rerank.rerank(query, papers, model_name=None).

How it works:

1. The aggregator fetches up to limit * 3 candidates (capped at 50).
2. Each candidate's text (title + abstract) is embedded alongside the query using the chosen SentenceTransformer model (cached per process).
3. Candidates are sorted by cosine similarity (descending).
4. Papers with neither a title nor an abstract keep their merge-order rank and are placed at the end.
5. The top --limit results are returned.

Identifier formats

paperhound accepts whatever you have on hand:

• arXiv ids: 2401.12345, 2401.12345v3, cs.AI/0301001, arXiv:2401.12345
• DOIs: 10.1038/s41586-020-2649-2, doi:10.1038/...
• Semantic Scholar paper ids: 40-char hex
• URLs: arxiv.org/abs/..., arxiv.org/pdf/..., doi.org/..., semanticscholar.org/paper/...

Configuration

Env var	Purpose
OPENALEX_MAILTO	Optional. Adds your email to OpenAlex requests so they land in the polite pool (better rate limits).
CROSSREF_MAILTO	Optional. Same idea for Crossref's polite pool.
CORE_API_KEY	Required to enable the CORE provider. Without a key the provider reports unavailable and the aggregator skips it silently. Get a free key at https://core.ac.uk/services/api.
SEMANTIC_SCHOLAR_API_KEY	Optional. Semantic Scholar's anonymous quota is shared globally and 429s are common; the provider retries with exponential backoff. Set this to your own key for steadier throughput.
PAPERHOUND_LIBRARY_DIR	Override the library directory (default ~/.paperhound/library/).

Adding a new provider

paperhound.search is a registry of provider factories. To add a new source:

1. Create src/paperhound/search/<name>.py with a class subclassing SearchProvider. Declare its capabilities (TEXT_SEARCH, ID_LOOKUP, OPEN_ACCESS_PDF) and override available() if it needs an API key.
2. Add unit tests in tests/unit/test_<name>.py that mock HTTP with respx.
3. Register it in src/paperhound/search/__init__.py with one register("name", Factory) call. Done — the CLI picks it up automatically.

Use it from agents

paperhound ships a ready-to-install skill at skills/paperhound/SKILL.md that documents every command, recommends the JSON output flag, and gives an end-to-end example. Install it with one command:

npx skills add alexfdez1010/paperhound

This uses the skills CLI to discover the SKILL.md under skills/paperhound/ and place it in your agent's skill directory (~/.claude/skills/paperhound/ for Claude Code). Pass -a <agent> to target a specific agent (e.g. -a claude-code, -a opencode).

Development

make install            # uv sync --extra dev
make test               # unit tests (network-free, respx-mocked)
make test-integration   # live API tests — always live, no env-var gate
make test-all           # unit + integration
make check              # lint + format check + unit tests (run before pushing)

Unit tests use respx to mock HTTP, so they never touch the network. Integration tests under tests/integration/ always hit the real provider APIs (arXiv, OpenAlex, DBLP, Crossref, Hugging Face Papers, Semantic Scholar) — no env-var gate, no mocks. The SemanticScholarProvider retries 429s with exponential backoff; export SEMANTIC_SCHOLAR_API_KEY only if you want faster runs.

Releasing to PyPI

1. Bump version in pyproject.toml.
2. Push to main. The Publish to PyPI workflow builds and publishes via PyPI Trusted Publishing — idempotent on the version field, so re-pushing the same version is a no-op.

License

MIT — see LICENSE.