Research Log

RAG (Retrieval-Augmented Generation) is the technique behind the "Ask Your Sources" feature. Instead of the AI making things up, it searches your actual PDFs for real evidence.

Indexing: When you first query an article, the system extracts the full text from your PDF and splits it into overlapping chunks (~500 words each), and also renders each page as an image. Both text and page images are sent to Google's gemini-embedding-2-preview model, which converts them into 3072-dimensional vectors — lists of 3072 numbers that represent the meaning of each passage or visual. These vectors and chunks are cached in Cloudflare KV so the extraction only happens once per article.

Searching: When you type a question, your query gets embedded into the same 3072-dimensional vector space. The system calculates the cosine similarity between your query vector and every chunk vector, ranking them by how close they are in meaning (not just matching keywords). It also searches page images to find relevant graphs, charts, and data visualizations. The top-scoring text chunks are sent to Gemma 3 12B, a 12-billion parameter language model running locally on an NVIDIA RTX 5070 Ti GPU, which reads those chunks and extracts key excerpts with page numbers.

AI feedback on your writing uses that same Gemma 3 model on the same GPU. The model runs through a Cloudflare Tunnel, so your writing never touches a third-party cloud — inference happens on hardware we own.

Your workflow: Ask a question → review the passages the AI found in your sources → write your own response explaining why the quote matters to your argument → insert it into your paper with an auto-generated APA citation.

Why it matters: This is the same retrieval pipeline used in professional research tools and production AI systems. You're learning to query a corpus, evaluate evidence, and build arguments — skills that transfer directly to tools like Semantic Scholar, Elicit, and LexisNexis.