Knowledge fabric: key concepts and content strategies

View summary

Key concepts of knowledge fabric

Chunking

Documents are split into smaller, manageable “chunks” (for example, paragraphs, sections). All ingested knowledge content is automatically divided into fixed-size chunks of 400 tokens each. This means that, regardless of the original document structure, the system splits the text into segments of up to 400 tokens (where a token is typically a word or part of a word). Each chunk is then indexed and used for search and answer generation.

Knowledge fabric uses fixed chunking because it ensures consistency and predictable performance across all documents.

Indexing 

Chunks are indexed and embeddings are generated. Each chunk has its own metadata that helps in traceability (for example, source document and URL).

Result ranking

The results are ranked using confidence scores.

• The top three chunks are returned by default, ranked by relevance/confidence.
• Each result includes a confidence score and source reference (name and URL)

Best practices to optimize and improve retrieval

While the chunking process is automated, the way you structure your source content can significantly impact the quality and relevance of search results and AI-generated answers.

Here are some best practices:

• Write in clear, self-contained sections: Ensure that each paragraph or section of your document can stand on its own. Avoid splitting important information across multiple paragraphs or pages.
• Use headings and subheadings: Organize your content with clear headings. This structure helps the chunking process capture logical breaks and improves the relevance of retrieved chunks.
• Avoid overly long paragraphs: Large blocks of text can split mid-sentence when chunked. Try to keep paragraphs concise and focused on a single topic.
• Summarize the key points: Place important information at the beginning of sections or paragraphs, so it is less likely to be missed if a chunk starts or ends mid-section.
• Tables and non-text content: Since only text is indexed, ensure that any critical information in tables or images is also described in the surrounding text.

Troubleshooting and other considerations

As chunking is based on token count, some chunks can start or end in the middle of a sentence. To minimize sentence fragmentation, keep the sentences and paragraphs concise.

• Retrieval quality: If you notice that search results are missing context or returning incomplete answers, review your source documents for opportunities to improve structure and clarity.
• Future enhancements: Genesys Cloud will introduce more advanced chunking strategies (such as semantic or heading-based chunking) in future releases, which will further improve retrieval quality.

Search limitations

• Minimum: Two characters (to support languages like Japanese)
• Maximum: 50 for answer generation

Currently, there is no way to add custom filters. You can organize your content in folders to ensure that only relevant content is available for selection in a knowledge configuration for a touchpoint.

Tips to improve answer generation

AI generates answers using the top retrieved chunks (a maximum of three) and the query, in the language in which the touchpoint is configured. All sources (chunks) used to generate the answer are cited with name and URL.

To improve answer generation, follow these best practices:

• Ensure good quality source: High-quality, well-structured source content leads to better answers.
• Chunk appropriately: Chunks must be context-rich and self-contained.