Agentic RAG for Drug Knowledge Retrieval, Reasoning, and Evidence Synthesis

Technical Report  ·  中文文档 / Chinese Version  ·  Full Resources

DrugClaw is a CLI and agent runtime for drug-focused questions. It prioritizes evidence-grounded retrieval, source attribution, and traceability over answers that only sound fluent.

• It is not a general chat assistant. It is built for real drug questions such as targets, indications, repurposing, safety, DDIs, PGx, and labeling.
• It aims to return structured evidence and source-grounded conclusions, not just a polished summary.
• It supports both a lightweight minimal mode and a deeper full local-resource mode.

If this is your first time using DrugClaw, follow the steps below and you should be able to run a real query quickly.

git clone https://github.com/QSong-github/DrugClaw
cd DrugClaw

pip install -e .

DrugClaw reads navigator_api_keys.json from the repository root by default. If you do not want to pass --key-file, put the config file at the repository root.

Create the file and provide at least these fields:

{
  "api_key": "<your-api-key>",
  "base_url": "<your-base-url>",
  "model": "gpt-5.4-mini"
}

If your config file lives somewhere else, pass it explicitly with --key-file <path>.

python -m drugclaw doctor

If the setup is valid, you should see something like:

Doctor result: setup looks usable.

python -m drugclaw run --query "What are the known drug targets of imatinib?"

If this works, you already have the minimal usable setup running.

• Drug targets and mechanisms
• Indications and repurposing evidence
• Safety risks and serious adverse reactions
• Drug-drug interactions
• Pharmacogenomics
• Labeling and prescribing information

python -m drugclaw run --query "What pharmacogenomic factors affect clopidogrel efficacy and safety?"
python -m drugclaw run --query "What are the clinically important drug-drug interactions of warfarin?"
python -m drugclaw list

Minimal mode is the default.

The repository only tracks a minimal resources_metadata/ subtree. That is enough for the CLI, basic queries, and the default test suite. For most new users, this is the right place to start.

Use minimal mode if you want to:

• try DrugClaw quickly
• run the default queries and tests
• avoid downloading large local resource bundles upfront

If you need deeper and broader local evidence coverage, download resources_metadata_full.tar.gz from the Hugging Face resource mirror, then extract it at the repository root:

tar -xzf resources_metadata_full.tar.gz

Recommended flow:

1. Download resources_metadata_full.tar.gz
2. Extract it at the repository root
3. Run python -m drugclaw doctor again

This is not just a generic data add-on. It expands the same resources_metadata/ tree in place, enables more LOCAL_FILE resources, and supports deeper local evidence retrieval.

Use full mode if you want to:

• increase local resource coverage
• enable more local-data-backed resources
• run deeper resource-level validation or analysis

• Show available resources and recommended entry points:

python -m drugclaw list

• Re-check your environment:

python -m drugclaw doctor

• Explore the built-in demo flow:

python -m drugclaw demo

• Repository guide: docs/repository-guide.md
• Maintainer guide: maintainers/README.md