Accelerating SOP Retrieval for Japanese Netsourcing Operations Using Machine Learning

Transforming Manual SOP Search into Instant Access with NLP and Vector Search

Client Persona 👤 ▶
A Japan-based enterprise operating in the netsourcing domain, this client specialises in:
  • Network infrastructure management
  • Device and traffic monitoring
  • System design, integration, hosting, and operations
Their core business involves delivering end-to-end IT infrastructure services to customers, ensuring high availability, performance, and security of networked systems. The client operates in a fast-paced environment where rapid access to accurate operational knowledge—such as Standard Operating Procedures (SOPs)—is critical for maintaining service quality and responsiveness.
Problem Statement ❓ ▶
The client maintained a vast repository of approximately 90,000 Standard Operating Procedures (SOPs) in Japanese on their internal wiki. These SOPs were critical for responding to customer issues and managing network operations. However, the manual search process was slow and inefficient, often taking hours to locate the correct SOP. This delay impacted response times and operational efficiency.
The client required a fast, reliable, and accurate retrieval system that could:
  • Automatically identify the most relevant SOPs based on case data
  • Deliver precise results in seconds
  • Support Japanese-language content and maintain high accuracy
Solution Overview 💡 ▶
To address the challenge of slow and manual SOP retrieval, we developed a Machine Learning-powered pipeline that delivers relevant SOPs in seconds based on input case data.
The system leverages Natural Language Processing (NLP) models trained on historical Japanese-language cases to:
  • Understand and interpret incoming queries
  • Retrieve the most relevant SOPs from a repository of over 90,000 documents
  • Automatically generate SOP responses tailored to the context
This solution significantly reduces manual effort and improves response time, enabling the client to deliver accurate and timely support to their customers.
Technical Architecture 🏗️ ▶
Component Technology / Tool
Language Python
Frameworks PyTorch, Langchain
Models BERT (fine-tuned for Japanese language understanding)
Embedding Model Multilingual embedding model (evaluated for Japanese compatibility)
Vector Database Weaviate (used for semantic search and structured metadata filtering)
Cloud Platform Google Cloud Platform (GCP)
Deployment Self-hosted on GCP with 2 GPUs; Weaviate hosted in a separate Docker container
Model Strategy Fine-tuned transformer models for Japanese text, integrated with Langchain for RAG
Prompt Orchestration Langchain-based wrapper for dynamic prompt engineering and response generation
Implementation Steps ✅ ▶
The solution was implemented in two core stages: Stage 1: Retrieval of Top 5 Relevant SOP Wiki Links
  • Developed a semantic search mechanism using vector embeddings to identify and present the top 5 most relevant SOP links from the client's wiki based on input case data.
Stage 2: SOP Generation Based on Problem Statement
  • Built a Machine Learning pipeline to generate SOPs dynamically using historical case data and fine-tuned LLMs.
Outcome & Impact 📊 ▶
The implementation of the machine learning-based SOP retrieval system delivered significant operational benefits:
  • SOP retrieval time reduced from hours to seconds Enabled rapid access to relevant procedures, drastically improving response times.
  • Increased consistency in issue resolution Standardised responses across teams by ensuring the right SOPs were always retrieved.
  • Improved operational efficiency and reduced manual effort Freed up valuable engineering and support resources by eliminating time-consuming manual searches.
Challenges & Learnings 🛠️ ▶
Implementing a high-performance SOP retrieval system for Japanese-language content presented several unique challenges and valuable insights:
  • 🧩 Japanese Language Complexity
    Handling linguistic nuances required specialised tokenisation and pretraining strategies. The limited availability of open-source LLMs trained on Japanese posed constraints in model selection and embedding accuracy.
  • 🧠 Model Accuracy Across Diverse SOP Formats
    SOPs varied widely in structure and terminology. Ensuring consistent and accurate retrieval demanded iterative feedback loops with domain experts to refine annotations and model behaviour.
  • 🖥️ Self-Hosting Infrastructure
    Deploying the solution in a self-hosted environment introduced challenges around latency, concurrency, and hardware optimisation. Recommending the right infrastructure setup to balance performance and cost was a critical part of the deployment strategy.
Future Scope 🔭 ▶
To further enhance the SOP retrieval system and extend its capabilities, the following initiatives are planned:
  • 🌐 Multilingual SOP Retrieval
    Expand support beyond Japanese to include other languages, enabling broader applicability across global operations.
  • 🔄 Feedback Loop Integration
    Implement continuous learning mechanisms by incorporating user feedback to improve model accuracy and relevance over time.
  • ⚠️ Anomaly Detection for SOP Gaps
    Introduce intelligent monitoring to identify missing or outdated SOPs, ensuring comprehensive coverage and proactive updates.