mathesis-ai 0.1.4

Interdisciplinary GraphRAG -- find cross-domain bridges in scientific literature

mathesis-ai

Find cross-domain bridges in scientific literature. Give it a research question; it returns papers from unrelated fields that solved the same underlying problem -- expressed in different vocabulary.

Built on SPECTER embeddings, recursive semantic graph expansion, PDF RAG, and Groq LLM summaries. No server required.

---

Install

# Install CPU torch first (PyPI's default is a 4 GB CUDA build)
pip install torch --index-url https://download.pytorch.org/whl/cpu

pip install mathesis-ai

GPU:

pip install torch --index-url https://download.pytorch.org/whl/cu121
pip install mathesis-ai

Requires Python 3.11+. Needs ~4 GB RAM for the embedding model.

---

Usage

from mathesis import Mathesis

m = Mathesis(groq_api_key="gsk_...")  # free key at console.groq.com
m.seed()                               # fetch & embed 250 papers, ~10 min, run once

results = m.query("vanishing gradients in deep neural network optimization")

for bridge in results.bridges:
    print(bridge.domain_jump)   # "q-bio.NC -> cs.LG"
    print(bridge.confidence)    # 0.796
    print(bridge.llm_summary)   # one-paragraph analogy explanation

No .env file, no server to start. Data is stored in ~/.mathesis by default.

---

Configuration

m = Mathesis(
    groq_api_key="gsk_...",          # required for LLM summaries
    data_dir="~/my-mathesis-data",   # default: ~/.mathesis
    domains=["cs.LG", "q-bio.NC"],   # default: 5 arXiv categories
    llm_enabled=True,                # default: True
    device="cpu",                    # "cpu" or "cuda"
)

---

Query response

results.bridges_found          # bool
results.bridges                # list of BridgeResult
results.top_nodes              # top 5 papers by graph rank
results.graph_stats            # {"nodes": 310, "edges": 1548, "domains": [...]}

bridge.domain_jump             # "q-bio.NC -> cs.LG"
bridge.confidence              # 0.0 - 1.0
bridge.similarity              # raw cosine similarity
bridge.rag_confirmed           # bool - methods text confirmed the bridge
bridge.rag_excerpt             # extracted methods section passage
bridge.llm_summary             # Groq-generated analogy explanation
bridge.source_paper.title
bridge.source_paper.id         # arXiv ID
bridge.target_paper.title

---

Async support

results = await m.aquery("attention mechanisms in transformers")

---

Status

print(m)
# Mathesis(papers_indexed=581, domains=5, llm=enabled)

m.status()
# {"papers_indexed": 581, "domains": [...], "data_dir": "~/.mathesis"}

---

Getting more data

m.seed(papers_per_domain=50)    # quick start -- 250 papers total
m.ingest()                      # full corpus -- 500 papers/domain, ~25 min

---

Source & docs

https://github.com/nathanpi8/mathesis-ai