arXiv Inspiration

中文文档

A DeepSeek Harness (DSH) plugin that turns arXiv into a personalized research-inspiration feed: on-demand surveys filtered by a local interest knowledge base + LLM ranking, a settings-page UI to review and give feedback, an accumulating paper library, and full-text deep-read notes.

Features

• On-demand surveys — pick weekdays (UTC+8) and arXiv categories, or run from a session with the arxiv_inspiration tool. Papers are fetched, coarse filtered by your KB, ranked in parallel LLM batches, and distilled into per-paper recommendations (中文推荐理由/💡启发/建议行动 + English agent context).
• Settings-page UI (arXiv Inspiration card): configure categories/weekdays/ model/push preferences, launch surveys, review results with bilingual title/abstract, and give 👍/📖/👎 feedback — no auto-push, you decide what reaches a conversation.
• Feedback loop — 👍 auto-adds an interest chunk; 👎 gently down-ranks similar titles in later surveys (≤25% score penalty, single id exclusion). Feedback also feeds profile evolution.
• Paper library — every fetched paper accumulates in ~/.dsh-arxiv-inspiration/papers.jsonl with scores, feedback, Chinese translations, notes, and deep-read state. Searchable (keyword + optional semantic), browsable in the UI, and queryable by agents via arxiv_knowledge_base tool actions (search_papers, get_paper, add_paper, …).
• Deep-read notes — recommended papers can be read full-text (alphaXiv PDF Q&A as preferred source, local PyMuPDF map-reduce as fallback). Full notes live in ~/.dsh-arxiv-inspiration/analyses/<id>.md and open in a standalone viewer window; cards show a 3–5 bullet Chinese summary. Agents can read and write notes (read_analysis / write_analysis) and attach timeline notes (add_note / list_notes).
• Profile evolution — LLM drafts a revised research-interest profile from your feedback history; you adopt/edit/discard it in the UI. The plugin never silently rewrites your profile.
• alphaXiv assist (optional) — semantic paper discovery beyond configured categories and cited full-text Q&A, via your existing alphaxiv MCP credentials. Both toggles can be disabled per survey.
• Session push — push a digest (user-facing, bilingual) or a paper list + prompt (agent-facing) into any conversation you choose.

Requirements

• DSH host (see dsh.engines in package.json).
• An LLM provider configured in DSH (the plugin uses the host llm service and your session model selection by default; a specific provider/model can be pinned in settings).
• Node.js >= 22.
• Optional, for PDF full-text extraction and semantic search: Python (below).

All data is stored under ~/.dsh-arxiv-inspiration/. Uninstalling the plugin does not delete it.

Python setup (optional, recommended)

PDF deep-read and semantic search use a small Python environment. Without it the plugin still works — surveys, ranking, library, and notes run fine; PDF deep-read is skipped and semantic search falls back to keyword matching.

The recommended way is uv:

# 1. Install uv (macOS / Linux)
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Create a venv inside the plugin data dir (the plugin also auto-detects
#    .venv inside its own install directory)
mkdir -p ~/.dsh-arxiv-inspiration
uv venv ~/.dsh-arxiv-inspiration/.venv

# 3. Install the two extras
uv pip install --python ~/.dsh-arxiv-inspiration/.venv/bin/python pymupdf fastembed

Notes:

• pymupdf powers PDF text extraction (scripts/pdf_parse.py).
• fastembed powers semantic search (scripts/kb_vec.py, bge-small-en-v1.5, ONNX, no torch). The model downloads on first use to ~/.dsh-arxiv-inspiration/models/.
• System Python 3.12 with PEP 668 (externally-managed environment) rejects direct pip install — that is why a venv via uv is recommended (python3 -m venv may lack ensurepip on macOS).
• The plugin resolves Python as: .venv/bin/python in its install directory first, then .venv/bin/python in ~/.dsh-arxiv-inspiration/, then plain python3 on PATH.

alphaXiv credentials (optional)

The optional alphaXiv assist (semantic discovery + PDF Q&A with page citations) talks to https://api.alphaxiv.org/mcp/v1 with a Bearer token. The plugin does not have a config field for it — it reuses the token from a locally installed alphaxiv MCP bundle, so the credential lives in exactly one place:

~/dsh-bundles/alphaxiv-mcp/cordis.patch.yml

The plugin scans that file for a header line matching Authorization: Bearer <token> and uses it as-is. If you already run the alphaxiv MCP bundle, the file exists and nothing to do. Otherwise create it manually with the same shape the bundle uses, e.g.:

- insert:
    - id: alphaxiv-mcp
      name: dsh-mcp-client
      config:
        mcpServers:
          alphaxiv:
            url: https://api.alphaxiv.org/mcp/v1
            headers:
              Authorization: Bearer YOUR_ALPHAXIV_TOKEN

Notes:

• Get a token from your alphaXiv account settings; calls to discovery/PDF Q&A consume your alphaXiv assistant quota.
• The token never leaves your machine; nothing is embedded in this plugin or its repo.
• No file / no token: both alphaXiv toggles degrade silently — surveys still run, deep-read falls back to local PyMuPDF extraction.
• Token cached for the host's lifetime; edit the file and restart the DSH host to pick up a new one.

Install

dsh plugin add Lxion0553/dsh-arxiv-inspiration

Then restart the DSH host. The settings page gains an "arXiv Inspiration" section; sessions gain the arxiv_inspiration and arxiv_knowledge_base tools (open a new session after install).

Configuration

Via the settings page (persisted to ~/.dsh-arxiv-inspiration/config.json): categories, weekdays, cross-list toggle, max papers per digest, LLM provider/model, push count & prompt, research context, alphaXiv toggles.

License

MIT