Deeplink’s RAG system searches for, extracts, and uses relevant text fragments to answer questions. The quality and structuring of the source content directly determine the relevance of the generated answers.
Golden rule: aim to write each paragraph so that it can be understood on its own, clearly and precisely. Avoid vague wording so that each extract remains relevant once isolated in RAG.
1. Structure content for efficient retrieval
- Clear, hierarchical headings (H1, H2, H3) to allow precise segmentation.
- Short paragraphs (3–5 sentences max) to facilitate chunking.
- Bullet lists to isolate key points.
- Relevant keywords in headings and at the start of paragraphs.
2. Write explicitly and self-sufficiently
- Each section should be understandable out of context.
- Don’t hesitate to restate important concepts rather than referring to them indirectly.
- Favor complete, affirmative sentences.
3. Promote precision and disambiguation
- Avoid vague wording (“a few,” “some,” “important”) without quantification.
- Provide figures, dates, and definitions whenever possible.
- Don’t hesitate to clarify technical or specific terms. Adding a brief definition can greatly help the AI, even after the paragraph in question.
4. Optimize keyword search & summaries
- Naturally repeat key terms and relevant synonyms.
- Use lexical variants to maximize semantic coverage.
- Explicitly name entities, places, products, and concepts instead of using unexplained abbreviations.
- For long documents, it can be helpful to add a summary or abstract right at the beginning.
5. Ensure the quality of information
- Verify the accuracy and timeliness of data.
- Cite sources when relevant.
- Avoid internal contradictions that could disrupt indexing.
6. Pitfalls to avoid
- Avoid nesting long lists within other lists.
- Avoid referencing other parts of a document (legal) via chapter number (cf. art. 2.3); the AI doesn’t understand them. It is preferable to rename these references (with a textual title) or remove them.
- If tables contain important information, it is recommended to turn each row/cell into a separate sentence.
By following these best practices, you ensure optimal structuring of your documents and more effective use of AI. This best practices document was designed to help you be more autonomous in creating, importing, and managing your content, ensuring increasingly relevant results.