BioManner: Practice of an Intelligent RAG System for Bioprocess Engineering

BioManner is an Agentic RAG system designed specifically for bioprocess engineering. It improves the reliability of large language models in professional fields through problem decomposition, multi-step reasoning, and fact-checking mechanisms. The project adopts a dual-track research approach of "Evaluation + Enhancement" (BioMeter evaluation framework + BioManner enhancement system), supports local deployment based on Ollama, and is suitable for engineers, R&D personnel, and other groups, providing a reference implementation template for AI applications in professional fields.

BioManner corresponds to the academic paper titled "BioMeter and BioManner: Evaluating and Enhancing Reliability of Large Language Models in Bioprocess Engineering", which includes two complementary dimensions:

• BioMeter: A systematic evaluation framework for testing the reliability of large language models in the bioprocess engineering field, including benchmark tests, metrics, and datasets;
• BioManner: An enhancement system that improves model reliability through RAG technology and Agentic workflows (core content of the open-source code repository). The dual-track design reflects the pragmatic attitude of "understanding the problem first before solving it".

Document Processing and Indexing

• Input PDF documents are extracted and cleaned using PyMuPDF, semantically chunked (maintaining semantic integrity + configurable overlap), vectorized by embedding models with cache support;

Query Understanding and Problem Decomposition

• Complex problems are split into sub-problems (e.g., optimizing dissolved oxygen control in penicillin fermentation is split into 4 sub-problems);

Retrieval and Re-ranking

• Semantic retrieval (cosine similarity) + neural re-ranking (fine-grained ranking with Transformer models);

Multi-step Reasoning and Answer Synthesis

• Sub-problems are retrieved and answered independently, then integrated into a coherent result;

Fact-checking and Quality Control

• After generating answers, automatic fact verification and logical consistency checks are performed.

BioManner supports full local deployment based on the Ollama framework, with advantages including:

• Data Privacy: Sensitive documents do not leave the internal network;
• Cost Control: Reduces long-term usage costs;
• Multi-Model Collaboration: Different models are configured for different tasks (default embedding model: qwen3:0.6b, default problem decomposition model: deepseek-r1:7b, etc.), balancing performance and cost.

• Bioprocess Engineers: Quickly access technical documents to obtain process parameters and troubleshooting suggestions;
• R&D Personnel: Integrate multi-source knowledge to understand parameter impacts during new process development;
• Technical Trainers: Use as an intelligent Q&A assistant to support new employee training;
• Quality Control Teams: Verify the accuracy of technical documents using fact-checking mechanisms.

• Complete Implementation of Agentic RAG: Demonstrates the full workflow from decomposition to reasoning and synthesis, suitable for complex professional fields;
• Multi-Model Collaboration Strategy: Uses models of different scales for different tasks, pragmatically controlling costs;
• Local Deployment Priority: Meets the needs of data-sensitive fields;
• Quality Control Awareness: Fact-checking mechanisms improve reliability, which can be extended to high-risk scenarios.

Limitations

• Domain Specificity: Problem decomposition strategies are difficult to directly migrate to other fields;
• Document Format Limitations: Mainly supports PDF, with limited support for Excel, CAD instructions, etc.;
• Evaluation Standards: Evaluation details in the open-source code need further exploration;

Future Outlook

• Support more document formats and data sources;
• Introduce multi-modal capabilities (processing equipment images, flowcharts);
• Integrate industrial IoT data to achieve real-time process monitoring;
• Domain adaptation to migrate to similar professional fields.