Y-BAT – Your Personal Bioinformatics Assistant & Trainer
Y-BAT helps students and researchers learn bioinformatics, plan analysis projects and run real analyses in one place, with evidence-based answers that show their sources. It covers many core command-line tools, from FastQC and BWA-MEM2 to GATK, STAR and Kraken2.
Bioinformatics is learned from scattered manuals, forum threads and tutorials that disagree across tool versions. A general chatbot answers confidently, invents command flags, and cannot tell you which manual a claim came from. Beginners need plain explanations, while practitioners need exact, version-correct syntax, and both need somewhere to try an analysis without installing a dozen tools.
• Learn. A tutor for any bioinformatics question: a plain-language explanation first, then technical detail, key terms and follow-up questions, with numbered sources. It searches approved scientific websites only when the indexed documentation does not cover the question, and labels general background knowledge as such.
• Projects. Each project has its own chat for exact command syntax, workflows, troubleshooting, tool comparisons and current releases, answered from official documentation, the tools' own --help output and release pages. A guided planner turns a few questions into an editable, versioned analysis plan, and notes can be typed or dictated.
• Analyze. Real FastQC quality control and sequence identification (BLAST+, MAFFT, MView, Sanger AB1 traces) run in a network-isolated Docker sandbox. An assistant explains the results using only facts from the run, and lookups that would send sample data to NCBI or an external model wait for the user's approval.
The frontend is Next.js 16 with React and TypeScript. The backend is FastAPI with four LangGraph agents: an agentic RAG pipeline for project chat, a knowledge agent for Learn, a coordinator that hands Analyze questions to specialists, and a human-approval graph that pauses with interrupt() until the user decides.
Retrieval is hybrid: SQLite FTS5/BM25 for exact flags and names plus ChromaDB vectors, fused with reciprocal-rank fusion and reranked. Answers come from best set of LLMs.
• Retrieval: two retrieval strategies were compared with RAGAS question benchmark with validated reference answers. Hybrid search won on validation (composite 0.72 vs 0.69) and with faithfulness 0.82 on the held-back test split and with no hallucinated flags at the ≥ 0.95 cut-off.
• Model choice: Qwen 3.8 27B, GPT-5.6 Luna and DeepSeek V4 Flash models were evaluated and scored by automatic checks and a blind judge from another model family. Qwen (0.96) and Luna (0.97) tied; DeepSeek scored 0.78 with 4 failed answers. Qwen became the default because it is free and Luna as the fallback.
Y-BAT is live at https://y-bat.duckdns.org. Create a free account with your email address. For help with your own project or a tailored course, use the “Get a custom quote” form on the home page. You are welcome to give your constructive feedback via the ‘send feedback’ form on the home page.
