Lab 03
Visual RAG
Ask how this site was built
Find answers in the project library and open the sources behind them.
Overview
The answer exists somewhere in a library, but finding the right passage takes work.
How this could help a business
The same pattern supports internal knowledge tools, document search and support assistants. A company version needs access controls, maintained source documents and evaluation of real questions.
What I built
- A document preparation and search pipeline with a real vector database.
- Browser-side question processing and optional local answer generation with clickable citations.
Live demo
Try a suggested question. Search works without downloading the large AI writer.
What I built
Implementation, decisions and evidence.
What Sebastian built
The application around the technology.
- A document preparation and search pipeline with a real vector database.
- Browser-side question processing and optional local answer generation with clickable citations.
- A separate private-PDF path that reads, searches and keeps the document in the browser.
One engineering decision
Show sources even without the large answer-writing download, so visitors can check the evidence themselves.
What to check yourself
Open a retrieved passage and compare it with the question. Try a private PDF; its text is not uploaded.
Tools and model credits. bge-small creates search representations, Qdrant searches the site library, and Qwen through WebLLM writes optional answers.
These are portfolio demonstrations. Explore professional experience separately, or discuss how the approach could fit your team.
Learn
Understand what the application is doing.
Open a source and check whether it actually answers your question. A close search match is a clue, not a guarantee.
Pause and predict
Does a real citation guarantee that an AI sentence is supported?
A learning question, not a score. Try this in the working demo.
Try changing something
Now make it your experiment.
Ask a question the project library cannot answer. Check what the sources actually contain rather than trusting fluent wording.
A citation can point to a real source without supporting the sentence. Retrieval quality is the part I am still improving: the committed benchmark is below my target, and because the match cutoff was set on the same question set, I treat those numbers as optimistic until a held-out split exists.
A few words, explained
- Citation
- A link to the source passage used to support an answer.
Data handling & current limitations
The AI writer is an optional large download. You can read the sources on their own.
Demo only. Each lab below explains which inputs stay on your device and which data is sent for server-backed features. Model files are downloaded and may remain in your browser cache. Hosting and security services also receive the network requests needed to serve the site.
Your question is embedded in this browser. Searching the site corpus sends the query vector to our search service. The question text and any PDF you add stay on this device.
A citation can point to a real source without supporting the sentence. Retrieval quality is the part I am still improving: the committed benchmark is below my target, and because the match cutoff was set on the same question set, I treat those numbers as optimistic until a held-out split exists.