RAG & Document Grounding

Advanced

Stop hallucinations with real data

01 The Concept

Retrieval-Augmented Generation (RAG): inject real documents into the prompt. The model reasons over your data instead of its training memory.

This eliminates hallucination on domain-specific questions, keeps answers current, and makes AI useful for internal knowledge no model was ever trained on.

02 Weak vs. Strong

EX 01Employment contract analysis

Analyze this contract. Answer ONLY from the text above — do not use general legal knowledge.

Non-compete: Is there one? Quote the exact clause. Then in plain English: geography, duration, industries covered.
IP assignment: Quote it. Does it cover inventions I create on my own time? Is there a carve-out for personal projects?
Notice period: What does each party give? Any difference for cause vs. without cause?
Unusual clauses: Anything more restrictive than standard that I should ask a lawyer about?

Format: quote the relevant text first, then plain-English interpretation. If something is NOT in the contract, write "Not specified in this contract" — never guess.

→ Why it works

Document isolated in XML. 'ONLY from the text above' prevents general legal knowledge that may not apply.

EX 02Meeting notes → action items

<meeting_notes> [Paste raw meeting notes or transcript here] </meeting_notes>

From these notes ONLY, extract:

Decisions made — each as one present-tense sentence. Only include things explicitly decided, not discussed.
Action items — for each: owner (exact name from notes), action (what they committed to), deadline (if mentioned), blocking dependency (if any)
Open questions — raised but not resolved. Who answers each?
Deferred topics — explicitly pushed to a future meeting.

Rules:

If an action item has no named owner, flag it as [OWNER NOT SPECIFIED] — don't assign one
Do not add anything not in the notes
If notes are ambiguous about whether something was decided vs. discussed, ask me

→ Why it works

Four extraction categories prevent mixing decisions with actions.

03 Key Points

01Paste the document first, then ask — don't hope the model knows it

02'Answer ONLY from the provided documents' is the hallucination kill switch

03Ask for citations: 'quote the exact sentence supporting each claim'

04For long docs: ask for relevant sections first, then answer

05RAG is the most impactful single-technique hallucination reducer

04 Model-Specific Notes

Best model for long-document RAG — 200K context, excellent citation tracking. Use tags for multi-document sessions.

05 For Your Role

Paste the thing first — email, article, contract, report. Then ask 'Based on the above...' This eliminates almost all hallucination.

Next: Cost Optimization & Token Economics →