This phase is about giving your AI agent the right knowledge to work with. By starting with high-quality content and setting clear tone and rules, you’ll ensure the agent gives accurate, trustworthy answers from day one.
Think of this phase as stocking the shelves before you open the store.
Steps
1. Prioritize what to add first
Decide which sources are most valuable to upload.
Why it matters: Starting with essential docs avoids clutter and helps the AI answer clearly.
Examples:
Help center or product documentation (top priority)
Developer documentation for technical depth
Useful website pages (FAQs, pricing, product descriptions)
Avoid: marketing pages, duplicates, or outdated files.
2. Organize your files
Break large manuals into smaller, topic-focused documents.
Why it matters: Well-organized files improve search accuracy and answers.
Tips:
Title files clearly (e.g., installation – windows)
Remove duplicates
Keep documents grouped by purpose (setup, billing, troubleshooting)
RAG Checklist (Knowledge Hygiene)
Follow these practices to keep answers accurate and up to date:
Chunk smartly: split docs into semantic units (procedures, steps). Aim for 400–800 tokens per chunk.
Use hybrid retrieval: combine embeddings + BM25 for better coverage.
Rerank results: sort retrieved chunks for relevance before passing to the AI.
Track version IDs: upload docs with version labels so outdated answers can be spotted.
Narrow facts in Q&A: store limits, quotas, or license rules as Q&A entries.
Deprecate old docs: remove or redirect outdated articles to avoid stale answers.
3. Upload your key sources
Add the selected files and pages into Ordemio’s Knowledge Base.
Why it matters: The AI can only answer from what you provide.
Processing takes minutes — status shows in the knowledge section.
Content is used as context, not to train the model.
Order of upload doesn’t matter.
Example view of Knowledge Base upload, where documents and sources are added.
4. Set tone and rules
Define how your AI should sound and when it should escalate.
Why it matters: Consistency builds trust and prevents awkward replies.
Provide at least 1 example for email and 1 for chat (friendly, formal, concise).
Configure the escalation rules you defined in Phase 1 (Settings → Escalation Rules).
Document any operational guidelines (e.g., hours of service, blocked topics).
Example settings page for defining communication tone and configuring escalation rules.
5. Test in the Playground
Ask real customer questions to check coverage.
Checklist of test cases:
Top 10 FAQs
5 common “how-to” steps
3 tricky edge cases
6. Expand knowledge gradually
Once core topics work, add the next category of knowledge requests.
Why it matters: A step-by-step rollout keeps quality high.
Add content for the next most frequent knowledge topic.
Update or remove outdated docs as you go.
Note: This step covers knowledge content only. Expansion into integrations, actions, and workflows happens later in Phase 5.
Best Practices / Tips
Start small with just the most valuable sources.
Use one “source of truth” per topic to avoid conflicts.
Provide short, concrete tone examples your AI can mimic.
Upload in manageable chunks, not all at once.
Test with real customer wording (not just internal terms).
Common Mistakes to Avoid
Uploading every file you have — noise hurts accuracy.
Including irrelevant marketing content.
Leaving giant PDFs unsplit (hurts retrieval).
Skipping tone and escalation setup.
Going live without Playground testing.
Cross-references
Knowledge Base – Add data sources (for uploading docs)
Retest in the Playground
✅ Expected outcome: Core knowledge sources are uploaded, tone and escalation rules configured.