Bilgi Getirme ve Dayandırma

Search

1. Semantic EmbeddingsSentences as vectors: meaning geometry for retrieval.1–2 sa
2. Similarity SearchNearest neighbors at scale: exact scan vs approximate indexes.1–2 sa
3. Lexical Retrieval: BM25The strong sparse baseline: term frequency with saturation and document-length normalisation.1–2 sa
4. Approximate Nearest Neighbor IndexesGraph and partition indexes (HNSW, IVF, product quantisation) that trade exactness for speed at scale.1–2 sa
5. Hybrid Retrieval and Rank FusionCombining lexical and dense retrievers and merging their rankings with reciprocal rank fusion.1–2 sa
6. Training Dense RetrieversBi-encoders trained contrastively on query-passage pairs, with hard negatives doing most of the work.1–2 sa

Evidence

1. ChunkingSplitting documents for retrieval: boundaries, overlap, and the trade-offs.1–2 sa
2. RetrievalQuery to evidence: scoring chunks and returning the top-k.1–2 sa
3. RerankingCheap first, precise second: the retrieve-then-rerank cascade.1–2 sa
4. Evaluating Retrieval-Augmented GenerationSeparating retrieval quality from generation quality: context recall, faithfulness, and answer correctness.1–2 sa

Grounding

1. Retrieval-Augmented GenerationConditioning on evidence: retrieved context changes what generation may say.1–2 sa
2. Grounding and CitationClaims pinned to sources: support mapping and its limits.1–2 sa
3. Query Transformation and Multi-Hop RetrievalRewriting, decomposing, and iterating queries when one retrieval step cannot find the evidence.1–2 sa