OpenBioLLM
An Open-Source Multi-Agent Framework for Genomic Question Answering
OpenBioLLM is a fully reproducible multi-agent framework that coordinates open-source LLMs with NCBI genomic APIs to deliver faster, more accurate, and interpretable genomic QA. The agent routing stack outperforms monolithic LLM setups while remaining easy to debug and extend.
🏛️ Core Architecture and Design
OpenBioLLM orchestrates four agent families—Router, Tool Agents, Evaluator, and Generator:
| Agent | Function Description |
|---|---|
| Router | Chooses the next agent to invoke using the query, history, and evaluator feedback. |
| Tool Agents | Specialized workers (Eutils, BLAST, Web Search) that connect to external databases and search tools. |
| Evaluator | Decides whether the evidence collected so far is sufficient or more tools are needed. |
| Generator | Synthesizes conclusions from all previous agents and produces the final answer. |
🚀 Key Enhancements and Performance Advantages
OpenBioLLM achieves significant improvements over monolithic LLM approaches across several dimensions:
1. Performance
The framework achieved higher accuracy on key genomic QA benchmarks, outperforming a monolithic LLM agent:
- GeneTuring Tasks: Achieved an average accuracy of 0.849.
- GeneHop Tasks: Achieved an average accuracy of 0.830.
2. Efficiency
The multi-agent, optimized tool routing design substantially improved execution speed:
- GeneTuring Latency Contrast: Reduced from $48.5\text{s}$ to $23.9\text{s}$.
- GeneHop Latency Contrast: Reduced from $92.4\text{s}$ to $55.0\text{s}$.
3. Key Insight
OpenBioLLM revealed that smaller specialized agents can outperform larger ones (“too smart” effect).
- Optimal Configuration: The experiments showed the 32B pipeline + 14B tools configuration yielded the best trade-off.
✨ Interpretability and Scalability
The multi-agent design provides crucial advantages lacking in monolithic models:
- Traceability: The entire workflow is traceable and debuggable via LangSmith.
- Scalability: The design allows for the easy addition of new tool agents to expand capabilities.