OpenGenes MCP Server

import json import pytest import os from pathlib import Path from dotenv import load_dotenv from just_agents import llm_options from just_agents.base_agent import BaseAgent from opengenes_mcp.server import OpenGenesMCP # Load environment load_dotenv(override=True) # Project paths PROJECT_ROOT = Path(__file__).parent.parent DATA_DIR = PROJECT_ROOT / "data" TEST_DIR = PROJECT_ROOT / "test" # Load judge prompt with open(TEST_DIR / "judge_prompt.txt", "r", encoding="utf-8") as f: JUDGE_PROMPT = f.read().strip() # Load system prompt for test agent - use comprehensive prompt that includes aging evidence guidance from opengenes_mcp.server import get_prompt_content # Get the base prompt from HuggingFace base_prompt = get_prompt_content().strip() # Add the comprehensive aging evidence guidance that's missing from HuggingFace prompt comprehensive_aging_guidance = """ ## CRITICAL: Comprehensive Aging Evidence Queries **COMPREHENSIVE AGING EVIDENCE REQUIREMENT**: When users ask about "evidence of aging", "link to aging/longevity", "aging associations", or similar questions for a gene, you MUST query ALL 4 tables for complete scientific evidence: 1. **gene_criteria** - Aging-related criteria classifications 2. **gene_hallmarks** - Biological hallmarks of aging 3. **lifespan_change** - Experimental lifespan effects 4. **longevity_associations** - Human population genetics data **Example question patterns that require comprehensive queries:** - "What evidence of the link between X and aging do you know?" - "Evidence of X and aging" - "X gene aging associations" - "Link between X and aging" - "What evidence links X to aging?" **For these questions, ALWAYS include data from ALL 4 tables, not just experimental data.** ## Additional Query Guidelines - When showing lifespan effects, include both mean and maximum changes when available - Order lifespan results by magnitude of effect - Use LIKE queries with wildcards for multi-value fields """ SYSTEM_PROMPT = base_prompt + comprehensive_aging_guidance + """ ## CRITICAL INSTRUCTION You are a database assistant. For EVERY question, you MUST: 1. Use the db_query tool to get data from the database 2. Include the SQL query you used in your final answer 3. Base your answer ONLY on data from the database ALWAYS use the db_query tool first before answering any question about genes, aging, or lifespan data. In your response, include the SQL query that you used to answer the question.""" # Load reference Q&A data with open(TEST_DIR / "test_qa.json", "r", encoding="utf-8") as f: QA_DATA = json.load(f) answers_model = { "model": "gemini/gemini-2.5-flash-preview-05-20", "temperature": 0.0 } judge_model = { "model": "gemini/gemini-2.5-flash-preview-05-20", "temperature": 0.0 } # Initialize agents opengenes_server = OpenGenesMCP() test_agent = BaseAgent( llm_options=answers_model, tools=[opengenes_server.db_query, opengenes_server.get_schema_info, opengenes_server.get_example_queries], system_prompt=SYSTEM_PROMPT ) judge_agent = BaseAgent( llm_options=judge_model, tools=[], system_prompt=JUDGE_PROMPT ) @pytest.mark.skipif( os.getenv("CI") in ("true", "1", "True") or os.getenv("GITHUB_ACTIONS") in ("true", "1", "True") or os.getenv("GITLAB_CI") in ("true", "1", "True") or os.getenv("JENKINS_URL") is not None, reason="Skipping expensive LLM tests in CI to save costs. Run locally with: pytest test/test_judge.py" ) @pytest.mark.parametrize("qa_item", QA_DATA, ids=[f"Q{i+1}" for i in range(len(QA_DATA))]) def test_question_with_judge(qa_item): """Test each question by generating an answer and evaluating it with the judge.""" question = qa_item["question"] reference_answer = qa_item["answer"] reference_sql = qa_item.get("sql", "") # Generate answer generated_answer = test_agent.query(question) # Judge evaluation judge_input = f""" QUESTION: {question} REFERENCE ANSWER: {reference_answer} REFERENCE SQL: {reference_sql} GENERATED ANSWER: {generated_answer} """ judge_result = judge_agent.query(judge_input).strip().upper() # Print for debugging print(f"\nQuestion: {question}") print(f"Generated: {generated_answer[:200]}...") print(f"Judge: {judge_result}") if "PASS" not in judge_result: print(f"\n=== JUDGE FAILED ===") print(f"Question: {question}") print(f"Reference Answer: {reference_answer}") print(f"Current Answer: {generated_answer}") print(f"===================") assert "PASS" in judge_result, f"Judge failed for question: {question}"