PV212 – SEMINAR TOPIC
Martin Kňažovič
1 / 6
Intro
• Thesis topic
• Last week: DND
• More thought
• Too many systems
• Too much time in non-AI development
• Commercial use
• AI Agents - why instead of what
2 / 6
Inquiry data extraction
• Why?
• Manual process
• 1000x / month
• Can be automated
• How?
• LLMs for data extraction
• Langchain
• What?
• Business inquiry extraction via LLM
3 / 6
Current state
• Solution already in development
• Prompt engineering (PE)
• Can be done better
PE Process
4 / 6
Ideas
• Retrieval-Augmented Generation (RAG)
• Products tend to get misclassified
• Goal: improve general accuracy
• Evaluation dataset
• Prompt augmentation with correct examples
• Examples selected based on similarity with inquiry
• Knowledge source scrapping
• Product parameters incomplete in inquiry
• Goal: improve detection of parameters
• Build knowledge source (KS)
• Table with products and parameters
• Scrap customer website
• Let LLM (Prod.) inquiry -> KS
5 / 6
Ideas cont.
• Multiple ways to success
• Do everything ?
• Let LLM to decide ?
UNCLE SAM WANT
YOUR OPINIONS
6 / 6