ArXiv Research Intelligence — MCP Server

Vectorless RAG over ArXiv — BM25 + RRF + contextual compression — as an MCP server for Claude.

What this is

An MCP server that gives Claude live tools to search ArXiv, download and index full papers, and answer research questions from your personal paper library — all without embeddings, GPU, or model downloads.

Related MCP server: arXiv Research MCP Server

Vectorless RAG — why and how

Most RAG pipelines embed every text chunk with a neural model, store vectors, and retrieve by cosine similarity. This works but has real costs: GPU or slow CPU inference, ~90MB model downloads, embedding drift, and poor handling of exact technical terms like LoRA, RLHF, KV-cache.

v2 replaces the embedding step entirely with Okapi BM25 — the same algorithm that powers Elasticsearch and academic search engines. Zero model downloads, instant indexing, and it handles exact terminology perfectly.

The full pipeline

Your question
     │
     ▼
Query expansion ── generates keyword variants to maximize BM25 recall
     │
     ▼
Multi-query BM25 ── each variant scored against all chunks independently
     │
     ▼
RRF fusion ── Reciprocal Rank Fusion merges ranked lists
              score = Σ 1/(60 + rank) — no score normalization needed
     │
     ▼
Contextual compression ── sentence-level BM25 extracts only the 2–3
                          sentences per chunk that answer the question
     │
     ▼
Cited context returned to Claude

v1 vs v2 comparison

No Claude Pro? Use it entirely from your terminal

Claude Pro is only needed for the Claude Desktop chat interface. The full pipeline — search, fetch, index, RAG query — runs locally with test_local.py and test_mcp.py.

Setup

git clone https://github.com/RatnamOjha/arvix-mcp-server
cd arvix-mcp-server
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

Anaconda users: always run with the full venv path to avoid conflicts: /path/to/arvix-mcp-server/.venv/bin/python3 test_local.py ...

Finding your paper ID

The ID is the number at the end of any ArXiv URL:

https://arxiv.org/abs/2301.07041
                   ^^^^^^^^^^^^^ this is your paper ID

All commands

Every flag has a short alias. Use whichever you prefer.

# Search ArXiv live
python3 test_local.py -s

# Fetch and index a paper (direct CDN download — no rate limiting)
python3 test_local.py -f -i 2301.07041

# Query your whole library — full output
python3 test_local.py -q -F

# Query one specific paper only
python3 test_local.py -q -p 2301.07041 -Q "what are the limitations?" -F

# Build a multi-paper library and query across all of them
python3 test_local.py -f -i 2301.07041
python3 test_local.py -f -i 2305.10601
python3 test_local.py -f -i 2005.11401
python3 test_local.py -q -Q "how do these papers approach retrieval?" -F

# See everything indexed
python3 test_local.py -l

# Run edge case tests
python3 test_local.py -e

# Run everything at once
python3 test_local.py -a

Everything persists at ~/.arxiv-mcp/ between runs — fetch once, query forever.

Enable LLM answer generation (free)

By default the pipeline shows retrieval context. To get a full cited LLM answer, set a free Groq key — no credit card needed.

# 1. Get a free key at https://console.groq.com (takes 60 seconds)
# 2. Set it
export GROQ_API_KEY=gsk_your_key_here

# 3. Run — you now get retrieval + generated answer
python3 test_local.py -q -p 2005.11401 -Q "what is the main contribution?" -F

Groq's free tier: 14,400 requests/day on llama-3.1-8b-instant. Anthropic (ANTHROPIC_API_KEY) works as a fallback if you prefer.

Test the MCP protocol directly (without Claude)

test_mcp.py spawns the server as a subprocess and sends it real JSON-RPC messages — byte-for-byte identical to what Claude Desktop sends. If this works, the Claude Desktop integration will work too.

# Test all tools
python3 test_mcp.py

# Test specific tools
python3 test_mcp.py --tool search_papers
python3 test_mcp.py --tool fetch_paper --arxiv-id 2005.11401
python3 test_mcp.py --tool query_library --question "what is BM25?" --paper 2005.11401
python3 test_mcp.py --tool list_library

With Claude Pro — MCP integration

Configure Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "arxiv-research": {
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["-m", "src.server"],
      "cwd": "/absolute/path/to/arvix-mcp-server"
    }
  }
}

Use the full absolute path to your .venv Python — avoids conflicts with Anaconda or system Python.

Restart Claude Desktop (Cmd+Q, then reopen). The 🔨 hammer icon will show 6 new tools.

Example prompts

Search for recent papers on speculative decoding
Fetch paper 2305.10601 and add it to my library
What does my library say about KV cache optimization?
What does paper 2507.07171 say about the evaluation? (uses arxiv_id_filter automatically)
Summarize paper 2305.10601
Show me my reading list

MCP Tools

The arxiv_id_filter parameter on query_library restricts search to one specific paper — Claude uses this automatically when you mention a paper ID in your question.

Known edge cases and how they're handled

ArXiv API rate limiting (HTTP 429)

The API allows ~1 request/3 seconds. If you run multiple searches quickly you'll hit a 429.

Fix: PDF download never rate-limits — we go directly to https://arxiv.org/pdf/{id} which is a CDN with no limit. Metadata (title, authors) is fetched separately and falls back gracefully if the API is unavailable. The paper is still fully indexed and queryable even with placeholder metadata.

Garbled text from LaTeX PDFs

Physics and math papers use LaTeX-rendered fonts that pdfplumber misreads, producing output like b e a h A b r ξ o a s α v t.

Fix: We use pymupdf (fitz) as the primary extractor — it reconstructs reading order from glyph positions and handles multi-column layouts correctly. A post-processing step filters lines where average token length < 1.8 characters (the statistical signature of scrambled LaTeX fonts). pdfplumber remains as a fallback if pymupdf is not installed.

Anaconda overriding the venv Python

If you see (base) and (.venv) in your prompt simultaneously, conda is winning.

Fix:

conda deactivate
source .venv/bin/activate
which python3  # must show .venv path

Yellow squiggles in VS Code

VS Code is using the wrong interpreter.

Fix: Ctrl+Shift+P → "Python: Select Interpreter" → choose .venv.

Wrong python being used for test scripts

Always use the full venv path:

/path/to/project/.venv/bin/python3 test_local.py --fetch --arxiv-id 2005.11401

Project structure

arxiv-mcp-server/
├── src/
│   ├── server.py           # MCP server — registers all 6 tools
│   ├── arxiv_client.py     # ArXiv search + direct PDF download/extraction
│   ├── vectorless_rag.py   # BM25 + RRF + contextual compression
│   └── reading_list.py     # JSON-backed reading list
├── web/
│   └── index.html          # Live landing page (GitHub Pages)
├── tests/
│   └── test_core.py        # pytest suite (runs on every push via CI)
├── test_local.py           # Standalone demo — no Claude needed
└── pyproject.toml

Tech stack

Zero ML dependencies for retrieval. No PyTorch, no sentence-transformers, no model downloads for the RAG pipeline.

Development

pip install -e ".[dev]"
pytest tests/ -v --asyncio-mode=auto
ruff check src/

Roadmap

• Qdrant/ChromaDB backend option for larger libraries

• LLM-powered query expansion (replaces rule-based fallback)

• Multi-user support with shared paper libraries

• Semantic Scholar + PubMed as additional sources

• Citation graph traversal - fetch papers that cite or are cited by a paper

License

MIT — built by Ratnam Ojha