You have seen the pipeline — from page extraction to cosine ranking and cited answers. This is the next step: an annotated bibliography to explore RAG at your own pace, grounded in the primary research and tools behind what the demos just showed.
Presenting or sharing this material? Put this page on your final slide: attendees scan the code and run every demo on their own phone. Zero downloads, no login — take it, reuse it, explore afterwards.
Real-time interactive 3D projection of English Wikipedia embeddings (Word2Vec 10k). Hands-on visualization of PCA, t-SNE, and UMAP clustering without writing code.
A web-native map that renders millions of high-dimensional document vectors in WebGL. Inspect density clusters, semantic voids, and outlier clauses across enterprise-scale text.
The foundation of production RAG. Standard BERT needs 65 hours to find the most similar sentence pair among 10,000 sentences; SBERT’s siamese networks do it in ~5 seconds.
The algorithmic engine inside Chroma, Qdrant, Milvus, Weaviate, and pgvector. In short: shortcuts that skip most of the library — only a few shelves get checked, so lookup drops from scanning everything O(N) to O(log N).
The paper that coined “RAG”. Instead of stuffing knowledge into model weights, it keeps knowledge in a searchable index the model reads at answer time — so answers can cite external sources.
Solves basic RAG’s weak spot: high-level synthesis questions (“top 3 themes across all policies”). Uses an LLM to build a knowledge graph and summarize community clusters hierarchically.
Replaces the single-shot retrieve→answer flow with a self-critiquing loop: if retrieved context is weak or ambiguous, the agent reformulates the query and searches again before answering.