How to maximise the accuracy of Ask Eolas - the AI Answer Engine for Healthcare

1. Uploading Guidelines or Source Documents

• Use well formatted guidelines. When using tables, create these using table tools in your editor rather than using the space bar or tab key for example. This way, they are more easily parsed by the system and less likely to produce errors.

• Use well structured guidelines. Try to use consistent headers to break down guidelines into main sections, subsections etc. This helps the system understand the hierarchy of content.
• Avoid mixed-scope uploads (e.g., don’t combine paediatric and adult guidance in one file). Keep populations separate per document.
• Ensure clarity in titles — include population and indication in the filename (e.g. “Adult Pneumonia Guidelines.pdf”).
• Keep guidelines up to date and remove duplicates or out of date versions.
• Avoid scanned images and documents. These are more likely to be parsed incorrectly, especially if they are of poor resolution.
• If the document mixes multiple conditions, segment it before upload where possible.

2. Phrasing Questions Clearly

• Specify the population: e.g. “adult with CKD stage 3”, “neonate with sepsis”.
• Specify the indication or context — “empirical antibiotics for community-acquired pneumonia” will give a more specific answer than “antibiotics pneumonia”.
• Don’t use prompts that invite reasoning beyond the sources (e.g., “What would you recommend?” or “What do you think the diagnosis is?”).

3. When Reading the AI’s Output

• Check the references — Always check the source documents as generative AI can produce hallucinations, in particular, cross check drug doses, contraindications, special populations (e.g. neonates, pregnant, breastfeeding, renal impairment etc). Ask Eolas will help with this by bringing users to the specific line of the specific document in one click.
• Verify scope — confirm that the population and condition match your query.

4. Handling “No Answer” Situations

• The model is prompted to return a ‘no answer’ response if it cannot find content that answers the question.
• Use this as a cue to:
  1. Check if your document set includes the right guideline.
  2. Upload or select a more specific source for that population or condition.
  3. Re-phrase your question (population + indication + context).

5. Managing Knowledge Sources

• Toggle knowledge sources appropriately — Use the paediatric/adult toggle to filter relevant content for your population. The system will only show content matching your selected mode.
• Be aware of source hierarchy — When multiple guidelines exist for the same condition, the system may present options from different sources e.g local vs national, or departmental vs hospital-wide. Always verify which source is most appropriate for your context.
• Check knowledge base scope — If using the Knowledge Hub alongside local guidelines, verify whether the AI is drawing from your organisation's information  or external sources like NICE/BNF.

6. Leveraging Visual Content

• Include context for images — When uploading guidelines with embedded images or diagrams, ensure surrounding text provides adequate context for the visual elements.

7. Quality Assurance Practices

• Review AI confidence indicators — As an admin, periodically review the metrics for the questions asked within your organisation.
• Cross-reference conflicting information — If you receive answers that seem inconsistent with your understanding, check if multiple contradictory guidelines exist in your knowledge base.
• Use the feedback mechanism — Report inaccurate or unhelpful responses to help improve the system's performance over time.

8. Organisational Best Practices

• Maintain clean knowledge bases — Regularly audit your uploaded content to remove outdated or duplicate guidelines that could cause conflicting responses.
• Coordinate with colleagues — Ensure multiple team members aren't uploading different versions of the same guideline, which can lead to inconsistent AI responses.