epistract

Epistract

Turn scientific literature into structured biomedical knowledge.

Paper: Beyond RAG: Domain-Specific Agentic Architecture for Biomedical Knowledge Graph Construction — technical report describing the architecture, evaluation across 6 drug discovery domains, and the evolution from GraphRAG to comprehension-based extraction.

Epistract reads drug discovery documents — PubMed papers, bioRxiv preprints, patent filings, clinical trial reports, FDA labels — and builds a knowledge graph that captures the entities and relationships a scientist cares about: compounds, targets, mechanisms, trials, biomarkers, pathways, and how they connect.

It runs as a Claude Code plugin. You point it at a folder of documents. It reads them with a scientist's understanding, extracts structured knowledge using a schema grounded in 40+ established biomedical ontologies, validates molecular identifiers with RDKit and Biopython, and produces an interactive graph you can explore in your browser.

The Name

From Greek episteme (ἐπιστήμη) — structured scientific knowledge, the highest form of knowledge in Aristotle's epistemological hierarchy — combined with extract. Episteme is not opinion or belief; it is knowledge grounded in evidence, demonstration, and systematic understanding. That is what this tool produces: not a bag of keywords, but a structured representation of how scientific concepts relate to each other, traceable back to the source text.

---

Quick Start

1. Install

Requires Claude Code and Python 3.11+.

Install from GitHub — run these inside any Claude Code session:

/plugin marketplace add usathyan/epistract
/plugin install epistract@epistract

Then restart Claude Code. The plugin is now available in all sessions.

For developers — clone locally and install as a dev marketplace:

git clone https://github.com/usathyan/epistract.git

/plugin marketplace add /path/to/epistract
/plugin install epistract@epistract

Restart Claude Code after installing.

Verify installation:

/epistract-setup

This installs sift-kg (the knowledge graph engine) and checks for optional molecular validation libraries (RDKit, Biopython).

Set up API keys — sift-kg uses LiteLLM for entity resolution during graph building. Set one of the following in your environment before running the pipeline:

# Option A: OpenRouter (recommended — one key, many models; used in development)
export OPENROUTER_API_KEY="your-key-here"

# Option B: OpenAI (SIFT_ prefix required — sift-kg forwards to LiteLLM)
export SIFT_OPENAI_API_KEY="your-key-here"

# Option C: Anthropic (SIFT_ prefix required)
export SIFT_ANTHROPIC_API_KEY="your-key-here"

# Option D: Google Gemini (SIFT_ prefix required)
export SIFT_GEMINI_API_KEY="your-key-here"

# Option E: AWS Bedrock (standard AWS credentials — LiteLLM reads directly)
export AWS_ACCESS_KEY_ID="your-key"
export AWS_SECRET_ACCESS_KEY="your-secret"
export AWS_REGION_NAME="us-east-1"

# Option F: Local models via Ollama (no API key needed)
# Just ensure ollama is running: ollama serve

Note: OpenRouter and AWS Bedrock keys are read directly by LiteLLM (no prefix). OpenAI, Anthropic, and Gemini keys use the SIFT_ prefix because sift-kg manages them through its config layer. You can also set SIFT_DEFAULT_MODEL to choose the model (e.g., openai/gpt-4o-mini, anthropic/claude-haiku, ollama/llama3.3).

This is separate from Claude Code's own API access, which is handled by your Claude Code subscription.

PubMed API key (optional) — if you plan to use /epistract-acquire to search and download articles from PubMed, an NCBI API key increases rate limits from 3 to 10 requests/second. The key is free.

To obtain one:

1. Create a free NCBI account at https://www.ncbi.nlm.nih.gov/account/
2. Sign in, go to Settings > API Key Management > Create an API Key
3. Set it in your environment:

export NCBI_API_KEY="your-key-here"

Note: The PubMed connector works without this key — it just runs at lower rate limits. For small corpora (<20 articles) you likely won't notice the difference. For larger acquisitions, the key prevents rate-limit pauses.

2. Ingest Documents

/epistract-ingest ./my-papers/