Prepare Your Source Content

Knowledge-Centered Service

Knowledge-Centered Service (KCS) is a set of industry-standard best practices for knowledge management. They help you create accurate and consistent articles, which in turn improves knowledge grounding. We list some key practices here, but you can learn more at the Consortium for Service Innovation.

There are four key pillars to prepare your content for use with AI models via data libraries. You want your content to be specific, organized, detailed, and accurate. Consider how to apply these best practices to your use case. That way, you can create a more effective knowledge base that amplifies the strengths of generative AI.

Specific

• Consider the audience of each article. Customers prefer short, solution-focused answers, while service representatives or developers need more technical depth. Adjust the tone, structure, and level of detail to match the identified audience’s needs.
• Segment content into articles that focus on isolated topics. AI systems retrieve information in fragments, so articles that combine loosely related subagents reduce coherence, dilute embeddings, and increase the risk of the model combining mismatched ideas. Ensure that each article resolves a clearly defined user intent, and avoid addressing multiple unrelated or partially related topics in a single piece.
• Research and choose article topics strategically. If you’re unsure what topics to cover, start by identifying the typical scenarios, questions, and problems your users face. This research ensures each article addresses real user needs, without unrelated or unnecessary information.
• Choose clear, standard terms for key concepts and stick with them. Define abbreviations and mark outdated terms as deprecated. Consistent terminology helps AI agents correctly link related concepts across articles and avoid generating responses that mix up ideas or misinterpret terms.

Organized

• Use a clear structural hierarchy. Like people, AI systems understand content more successfully when it's logically structured. Use heading tags (H1–H6) to reflect the informational hierarchy with paragraphs, lists, and headings to signal conceptual boundaries and relationships. Clear information architecture, and sections that group semantically related ideas, improves overall chunking, parsing, and vectorization for AI consumption.
• Spread out long-form content across fields. Don’t store large bodies of information in a single text field. Break up content into meaningful fields that correspond to parts of the text, such as Question, Description, Resolution, and Exceptions. Explicit boundaries help to build a more precise search index and supports accurate retrieval.
• Separate content by audience using both pre-retrieval and run-time enforcement. The information intended for customers, service representatives, and developers should differ in depth, tone, and sensitivity. To prevent exposure errors and improve AI content retrieval:
  • For pre-retrieval separation, create dedicated fields or articles for each audience. Explicitly indicating a target audience helps agents minimize cross-audience ambiguity.
  • For run-time filtering, use access-tier metadata to ensure that agents retrieve only the content that’s appropriate for the user querying the system or the use case at hand.
• Use metadata to boost content retrieval. In RAG systems, queries retrieve relevant content segments before the model generates a final response. Metadata, like product tags, target audience, versioning, and access tiers, enables accurate filtering, ranking, and compliance, improving retrieval precision and ensuring agent responses are well-grounded.

Detailed

• Provide thorough, comprehensive information. Generative AI performs best with detailed, comprehensive information that it can synthesize and deploy at multiple levels of detail. Prioritize clear, complete explanations over word count, especially for technical content. Detail isn’t about length but explanatory sufficiency.
• Define the context, conditions, and assumptions for each step. Include version applicability, environmental variables, exception handling, and clarify the reasoning behind each step. Without this, models may fill gaps with plausible but incorrect information (i.e. hallucinations).
• Include real-world scenarios to enrich contextual grounding. Describe examples in detail, and include best practices or considerations for common use cases. Likewise, describe common user mistakes or high-frequency escalation paths so Al systems can respond with nuance rather than generic summaries.
• Annotate visual content like images, videos, and screenshots with clear alt-text and descriptive captions. While visuals are easy for people to understand, AI systems rely on structured descriptions to interpret and use that context correctly. If visuals are important to your knowledge base, use a parser built for image or multimodal processing. Agentforce Data Library detects your content modes and applies the most appropriate parser for you. Even so, adding specific details like objects, visible text, layout, colors, relationships, and overall context helps improve accuracy, searchability, and consistency across systems.

Accurate

• Check facts and existing articles before publishing. Make sure instructions, policies, and technical steps match an official source. Or build a subject matter expert (SME) review into your publishing workflow. If the information is wrong in the knowledge base, agents may repeat it confidently. And, if related articles say different things, you may get blended or internally inconsistent answers. While fact checking a new article, review related content for alignment or redundancy. Conflicting or redundant information creates content noise and confuses LLMs.
• Test agent behavior and knowledge retrieval before and after deployment. Before going live, test your agent in a low-stakes environment to confirm it retrieves and applies knowledge correctly. Tools like the Retriever Playground and Agentforce Testing Center enable you to generate test cases from the agent’s connected knowledge sources to evaluate grounding and response quality. After deployment, monitor ongoing retrieval quality by using Knowledge/RAG Quality Data and Metrics.
• Update content when systems change and clearly mark new versions. When products, policies, or regulations change, review the related knowledge immediately. Or build content updates into your development cycle. Accuracy requires ongoing updates, not one-time reviews. State which versions, dates, or conditions apply. Mark outdated processes so an agent doesn’t present old guidance as current.
• Use AI output and user feedback as a quality signal. Monitor how your agents perform in real sessions to identify outdated, incomplete, or misapplied guidance. Tools like Knowledge/RAG Quality Data and Metrics, Agent Optimization, and Agent Analytics enable you to inspect sessions, track every interaction, and analyze trends. Patterns in agent behavior, user corrections, or escalations can reveal hidden gaps in your knowledge base that can be addressed to improve accuracy and alignment.