RAG & Retrieval

Retrieval-augmented generation, vector search, chunking, reranking.

Splitting long documents into smaller passages before embedding and indexing them, since retrieval works on chunk-sized units.

e.g. Splitting a 50-page PDF into 500-token chunks.

Anchoring a model's answer in retrieved or provided source material so it is less likely to hallucinate.

e.g. Requiring the model to quote the exact source passage it used.

A second, more precise scoring pass over an initial set of retrieved candidates to reorder them by true relevance before they reach the model.

e.g. A cross-encoder reranker re-scores the top 50 vector-search hits.

Fetching relevant external documents at query time and feeding them into the model's context so it can answer with up-to-date or private information.

e.g. Searching a company wiki before answering an internal question.

A datastore optimized for storing embeddings and running fast nearest-neighbor similarity search at scale.

e.g. Pinecone, Weaviate, and pgvector are all vector databases.

Finding the most relevant documents by comparing embedding similarity rather than exact keyword matches.

e.g. A query about "car trouble" also matches a document about "vehicle breakdown".