Build Your Personal Document Assistant
How AI answers from data it was never trained on. One insurance policy, one question — see the whole pipeline from PDF to trusted answer, using 5 sample pages you can inspect. All demos run on your phone, no download.
One Flow — Tap Any Step to Explore
Start with Ingest, follow the arrows left to right — each pill explains itself
We turn text into Meaning Fingerprints (Vectors), store them, find the closest match to your question, and let the AI write the answer. You are here under each step explains it in one line — no deck needed. New here? Tap Ingest and keep going right. Only the last step (Answer) uses an LLM — everything before it is fast, deterministic machinery.
Where RAG Is Used — Generic
Same pattern everywhere
Customer Support
Manuals + tickets → cited answer. Faster resolution.
Healthcare
Records → private assistant for doctors (on-prem).
Finance / Insurance
Policies, claims → “what is covered?” with page ref.
Legal
Contracts → “find risky clause + page”.
Education
Textbooks → Q&A with sources for students.
Enterprise
Handbooks → “leave policy?” in seconds.
Beyond Basic RAG — When to Level Up
Each level trades cost for capability • all need re-processing on update
Basic RAG
Split → embed → Top-K → answer.
On update: re-chunk + re-embed only that file. Cheapest, fastest.
GraphRAG
Builds entity graph (Hospital ↔ covers ↔ City).
On update: rebuild graph. Great for “how are X and Y connected?” but heavy.
Agentic RAG
Agent loops: search → check → search again.
On update: no storage, but slower + higher token cost. For multi-hop questions.
All 3 require re-processing on change — Basic cheapest per file. See resources.html for one link per card.
Go Further — After the Session
Open resources.html — curated links, each matched to the concept it explains: TensorFlow Projector (real enwiki vectors), Embedding Atlas, Sentence-Transformers, Vector DB intro, RAG/GraphRAG/Agentic guides. QR code is on that page for the final slide.
uiyuvi.github.io/rag-explainer/resources.html