e9fc99ddc6
Vector search with cross-encoder re-ranking, hybrid BM25+vector retrieval, incremental index updates, and multiple LLM backends (Ollama local, OpenAI API).
13 lines
576 B
Markdown
13 lines
576 B
Markdown
|
|
|
|
Simple query in ChatGPT produced
|
|
|
|
Metric | Best For | Type | Notes
|
|
-- | -- | -- | --
|
|
Cosine Similarity | L2-normalized vectors | Similarity | Scale-invariant
|
|
Dot Product | Transformer embeddings | Similarity | Fast, especially on GPUs
|
|
Euclidean Distance | Raw vectors with meaningful norms | Distance | Sensitive to scale
|
|
Jaccard | Sparse binary or set-based data | Similarity | Discrete features
|
|
Soft Cosine | Sparse with semantic overlap | Similarity | Better for text-term overlap
|
|
Learned Similarity | Fine-tuned deep models | Varies | Best accuracy, slowest retrieval
|
|
|