Small language models (SLMs) offer efficient solutions for tasks like customer support and documentation. Their effectiveness relies on a well-structured knowledge management pipeline designed as a continuous loop. This article outlines the process, starting with knowledge management, moving through curation, publishing, user interaction, and looping back with feedback.

The process begins with a centralized knowledge repository that consolidates data from various sources, such as product manuals, support tickets, internal wikis, chat logs, and external inputs. Use Markdown for its compatibility with AI workflows and version control systems like Git. Platforms like Contentful can manage this content, enabling collaboration. This repository acts as a dynamic hub, ensuring all data is organized and accessible for the next stages.

Next, curate and refine the consolidated data. Categorize content into stable material (e.g., product specifications) for SLM training and dynamic content (e.g., FAQs) for retrieval-augmented generation (RAG). Use large language models (LLMs) to draft and structure content, such as turning support tickets into guides. Apply manual guardrails—human oversight to ensure accuracy, clarity, and alignment with organizational standards. This dual-layer refinement produces high-quality content ready for publishing.

Once refined, publish the content for three key purposes:

1. Train SLMs: Use stable, curated content to fine-tune SLMs, enabling them to handle domain-specific tasks effectively.
2. Create RAG Content: Store dynamic content in a RAG system for real-time retrieval, ensuring SLMs can access up-to-date information.
3. Generate Documentation: Produce user-facing materials like manuals and FAQs, ensuring consistency across all outputs.

This publishing step ensures content is distributed efficiently to support both AI models and user needs.

The published content is then used by SLMs and RAG systems to serve users, answering queries and providing documentation. Track all user interactions to gather feedback, such as whether responses were helpful or if follow-up questions arose. This feedback identifies gaps, like outdated RAG data or misunderstood terms, and feeds directly back into the curation and refinement stage, completing the loop.

This continuous loop transforms SLMs into specialized tools that adapt to evolving needs. For example, a retail SLM can streamline customer support by accessing the latest return policies via RAG, while its training data ensures consistent responses. The feedback loop keeps the system current, enhancing reliability and user satisfaction.

A knowledge management pipeline for SLMs is a dynamic, cyclical process that integrates data collection, curation, publishing, and feedback. Start by centralizing your data in Markdown, experiment with LLM-driven curation, and establish feedback mechanisms. This approach ensures your SLMs deliver precise, scalable solutions tailored to your organization.