RAG Proxy, TechKnowFile 2026
University of Toronto · May 2026. Long deck. Picture first, notes underneath.
Failed fine-tuning runs are also told as a single page: Six failed runs.
Recording
May 2026 · TechKnowFile
Slide 1 of 18
Title
Open slide →
Slide 2 of 18
The problem
Open slide →
Slide 3 of 18
What is fine-tuning
Open slide →
Slide 4 of 18
Run 1: no system prompt
Open slide →
Slide 5 of 18
Run 2: question bloat
Open slide →
Slide 6 of 18
Run 3: quantization
Open slide →
Slide 7 of 18
Run 4: fabricated email
Open slide →
Slide 8 of 18
Run 5: RAFT
Open slide →
Slide 9 of 18
Run 6: the scenario
Open slide →
Slide 10 of 18
The research
Open slide →
Slide 11 of 18
Retrieve, augment, generate
Open slide →
Slide 12 of 18
Without the proxy
Open slide →
Slide 13 of 18
With the proxy
Open slide →
Slide 14 of 18
Technology stack
Open slide →
Slide 15 of 18
Security and privacy
Open slide →
Slide 16 of 18
Live demo
Open slide →
Slide 17 of 18
More features
Open slide →
Slide 18 of 18
Questions
Open slide →