An MCP server that gives Claude semantic paper search, full RAG over your library, and a vectorless retrieval pipeline with no embeddings, no GPU, and no paid subscription to get started.
Live ArXiv search. Returns ranked results with titles, authors, abstracts, and PDF links. Supports category filters like cs.LG, cs.CL.
Downloads PDF directly from ArXiv CDN (no rate limiting), extracts text with pymupdf, chunks at 500 words, indexes into BM25. Metadata fetched separately with graceful fallback.
Full vectorless RAG: Groq query expansion into 3 BM25 variants, RRF fusion, sentence-level compression, cited context. Use arxiv_id_filter to scope to one paper.
Pulls intro, methods, and conclusion chunks. Returns structured context for problem, contribution, results, and limitations.
Lists all papers indexed in your local BM25 store with titles, authors, categories, and chunk counts.
Your saved reading list with metadata, timestamps, read/unread status, and personal notes, persisted to JSON.
Without Claude Pro -- the full ML pipeline runs locally from your terminal.
test_local.py runs search, fetch, indexing, RAG queries, and LLM answer generation.
test_mcp.py tests the actual MCP protocol stack, byte-for-byte identical to Claude Desktop.
With Claude Pro -- connect to Claude Desktop and research in natural language.
Claude calls the 6 tools automatically based on what you ask.
Get a free Groq API key at console.groq.com (60 seconds, no credit card). Set GROQ_API_KEY and the server automatically uses llama-3.1-8b-instant for query expansion and answer generation. Anthropic works as a fallback. Nothing breaks if neither key is set -- retrieval still works fully.
The ID is the number at the end of any ArXiv URL:
https://arxiv.org/abs/2301.07041
Version suffixes like v2 are stripped automatically.
Every long flag has a short version. Use whichever you prefer.
Spawns the MCP server as a real subprocess and sends JSON-RPC messages over stdin/stdout -- byte-for-byte identical to what Claude Desktop sends. Handles the initialize handshake, tool listing, and tool calls. If this passes, Claude Desktop integration will work. It is the closest you can get to testing the full MCP stack without a Pro subscription.
The ArXiv metadata API allows roughly 1 request per 3 seconds. Burst fetching triggers HTTP 429 errors that block all requests.
Physics and math papers use LaTeX-rendered fonts that pdfplumber misreads, producing output like: b e a h A b r o a s instead of real sentences.
Seeing both (base) and (.venv) in your prompt means conda is winning and the wrong Python is running.
Default view shows 600 characters. Long papers or multi-source queries get cut off. Truncated context is never useful.
Run all 6 edge cases with one command to verify your setup handles every known failure mode.
Traditional RAG embeds every chunk with a neural model and retrieves by cosine similarity. Vectorless RAG replaces the embedding step with BM25 keyword retrieval -- the same algorithm powering Elasticsearch and Google Scholar. It handles exact technical terms like LoRA, RLHF, KV-cache that dense vectors often miss. Query expansion via Groq (free) generates 3 BM25 variants. Results are fused with Reciprocal Rank Fusion (RRF), then each chunk is compressed to only the relevant sentences. Final answer generated by Groq.
Claude Desktop connects over stdio via MCP. Research questions route to the right tool as structured JSON-RPC.
PDFs fetched from arxiv.org/pdf/{id} -- ArXiv CDN with no rate limit. pymupdf extracts text and handles LaTeX multi-column layouts.
500-word chunks with 50-word overlap. Each chunk added to Okapi BM25 instantly. No encode() call, no model load, indexing in under 1 second per paper.
Groq expands the query into 3 variants. Three BM25 searches fused by RRF. Optional arxiv_id_filter scopes search to one paper.
Compressed chunks formatted as [Source N] citations. Groq generates a grounded answer. Claude presents it in conversation with full attribution.