PaperHub

PaperHub fetches HuggingFace Daily Papers by programmatic date filters, assigns each paper to its own AI summarization agent, and renders Jupyter-friendly summaries by default.

Two ways to use PaperHub:

1. Terminal CLI (paperhub) — interactive launcher with a REPL interface. Primary mode.
2. Python / Jupyter — import PaperHub and call hub.run(...) directly.

Install

pip install paperhub

Optional provider extras:

pip install "paperhub[anthropic]"  # adds the Anthropic client
pip install "paperhub[google]"     # adds the Google Gemini client

For local models via Ollama, no extra package is needed — install Ollama and pull the model before use:

ollama pull gemma4:e2b   # ~2 B parameters, fast on CPU
ollama pull gemma4:e4b   # ~4 B parameters, higher quality

Configuration

PaperHub reads configuration from shell environment variables and from its own per-user config file. The CLI can save provider keys for you without touching a project-level .env file:

paperhub set-key openai
paperhub set-key anthropic
paperhub set-key google
paperhub check-llm
paperhub check-llm ollama     # no key needed — checks local Ollama server
paperhub api-keys
paperhub config-path

Inside the interactive launcher, use /set-key openai. The config file lives under the OS-specific user config directory, for example ~/Library/Application Support/paperhub/.env on macOS. Environment variables still override values saved there.

After set-key, PaperHub immediately sends one tiny request to the selected provider/model and reports whether the key and LLM are working. You can repeat that check later with paperhub check-llm or /check-llm.

Terminal CLI (Primary)

Start the interactive launcher:

paperhub

The launcher opens a REPL with a status dashboard showing the current provider, model, API key state, date range, and top paper count. Use commands to configure and run:

/help
/status
/version
/provider
/provider openai
/provider ollama
/model
/model gpt-5.4-mini
/model gpt-4.1-mini
/model default
/language
/language tr
/date 2026-05
/date 2026-05-15
/date 2026-W18
/date 2026-05-01 2026-05-31
/top 5
/concurrency 2
/metadata
/run
/set-key openai
/keys
/check-llm
/check-llm ollama
/config-path
/guide
/api-keys
/clear-cache
/clear-cache summaries
/clear-cache pdfs
/clear-cache keys openai
/clear
/quit

Date formats for /date

/metadata fetches HuggingFace paper metadata only and does not call an LLM. If the selected provider key is missing, /run prints setup guidance. /api-keys shows key status and setup help. /check-llm sends a tiny live provider request and confirms that the selected key/model can respond. /clear-cache deletes cached summaries/PDFs, and /clear-cache keys openai removes a saved cloud-provider key from PaperHub's user config.

CLI startup commands are available without entering the launcher:

paperhub version
paperhub set-key openai
paperhub keys
paperhub check-llm
paperhub check-llm ollama
paperhub api-keys
paperhub config-path
paperhub clear-cache summaries
paperhub clear-cache keys openai

You can also pass startup flags:

paperhub --provider anthropic
paperhub --model claude-sonnet-4-6
paperhub --provider ollama --model gemma4:e2b
paperhub --language tr
paperhub --top-n 10
paperhub --concurrency 2

Python / Jupyter API

from datetime import date
from paperhub import PaperHub

# OpenAI — default reasoning model
hub = PaperHub(provider="openai")
hub.run(period="month", year=2026, month=5, top_n=10)

# OpenAI — budget option (standard chat pricing, no reasoning tokens)
hub = PaperHub(provider="openai", model="gpt-4.1-mini")
hub.run(period="month", year=2026, month=5, top_n=10)

# Local model via Ollama (no API key required)
hub = PaperHub(provider="ollama", model="gemma4:e2b")
hub.run(period="month", year=2026, month=5, top_n=5)

# Single day
hub.run(period="day", year=2026, month=5, day=1, top_n=5)

# ISO week
hub.run(period="week", year=2026, week=18, top_n=5)

# Full year
hub.run(period="year", year=2026, top_n=20)

# Custom range
hub.run(period="custom", start=date(2026, 4, 15), end=date(2026, 4, 30), top_n=15)

run returns a list[PaperSummary]; in non-Jupyter contexts pass display=False and call render_plain yourself if you do not need the Markdown side effect.

For a step-by-step notebook, open examples/03_jupyter_quickstart.ipynb.

Provider Selection

PaperHub(provider="openai")                             # OpenAI default model
PaperHub(provider="openai", model="gpt-5.4-mini")       # reasoning model (default)
PaperHub(provider="openai", model="gpt-4.1-mini")       # budget: no reasoning tokens
PaperHub(provider="anthropic")                          # Anthropic default model
PaperHub(provider="anthropic", model="claude-sonnet-4-6")
PaperHub(provider="google", model="gemini-3-flash-preview")
PaperHub(provider="ollama", model="gemma4:e2b")         # local — no API key needed
PaperHub(provider="ollama", model="gemma4:e4b")         # local — higher quality

If model is omitted, PaperHub picks the selected provider's default model. Provider-specific config values such as PAPERHUB_OPENAI_MODEL override those defaults. PAPERHUB_MODEL remains available as a global override. OpenAI uses PAPERHUB_OPENAI_REASONING_EFFORT=medium by default; set it to an empty value to let the OpenAI API choose its model default.

Provider SDKs are imported lazily — installing paperhub includes the OpenAI SDK by default. Anthropic and Google require their optional extras. Ollama uses its native local HTTP API first and falls back to Ollama's OpenAI-compatible endpoint, so no extra Python package is needed beyond the base install.

OpenAI model guide

Local models via Ollama

PaperHub connects to Ollama's OpenAI-compatible endpoint at http://localhost:11434/v1 by default. No API key is required.

# Install Ollama: https://ollama.com
ollama pull gemma4:e2b    # ~2 B params — good for CPU inference
ollama pull gemma4:e4b    # ~4 B params — better quality

# Verify before running PaperHub:
paperhub check-llm ollama

Override the server URL if Ollama runs on a different host:

export PAPERHUB_OLLAMA_BASE_URL=http://192.168.1.10:11434/v1

Caching

PaperHub caches metadata, PDF text, and summaries in the OS-specific user cache directory, for example ~/Library/Caches/paperhub/paperhub.sqlite on macOS and ~/.cache/paperhub/paperhub.sqlite on Linux. Override with PAPERHUB_CACHE_DIR. Summary entries are keyed by (arxiv_id, model, language), so swapping models or output language gives you a clean re-run while keeping the PDF download free.

A second invocation with the same papers and model:

• Reuses the cached PDF text (no arXiv hit, no extraction).
• Reuses the cached summary (no LLM call).

Tests, Lint, Typecheck, Build

python3 -m pytest          # unit tests, no live network or LLM keys needed
python3 -m ruff check .
python3 -m ruff format --check .
python3 -m mypy src tests
python3 -m build           # build wheel + sdist

The unit tests use a fake LLM client and httpx.MockTransport; no real provider keys are required.

Project Layout

src/paperhub/
  __init__.py           # PaperHub public API
  config.py             # Settings (pydantic-settings)
  models.py             # PaperMeta, PaperSummary, RunRequest
  dates.py              # period → (start, end)
  fetchers/             # HF JSON API (default) + HTML fallback
  pdf/                  # arXiv download + text extraction
  agents/               # LLMClient protocol + provider clients + PaperAgent
    openai_client.py    # OpenAI (default, included)
    anthropic_client.py # Anthropic (optional extra)
    google_client.py    # Google Gemini (optional extra)
    ollama_client.py    # Ollama local models (no extra needed)
  orchestrator.py       # asyncio.Semaphore parallelism
  cache.py              # SQLite cache
  formatter.py          # Markdown / plain-text rendering
  interactive_cli.py    # `paperhub` interactive launcher
tests/                  # pytest suite, mocked HTTP and fake LLM
docs/ARCHITECTURE.md
docs/HOW_TO_START.md
docs/API_KEYS.md
examples/README.md
examples/03_jupyter_quickstart.ipynb

Troubleshooting

• "No papers found": HuggingFace may not yet have published Daily Papers

for that date. Use /metadata in the interactive launcher to check the fetcher without an LLM call.

• PDF text comes back tiny: some arXiv PDFs use unusual layouts. PaperHub

falls back to pdfplumber; if both extractors are short, the agent will pass through the abstract as the input text instead of failing.

• Ollama connection error: make sure Ollama is running (ollama serve) and

the model is pulled (ollama pull gemma4:e2b). Run paperhub check-llm ollama to verify before starting a full run.