Please enable JavaScript.

Coggle requires JavaScript to display documents.

AI-103/200 - Coggle Diagram

- - - - cosine similarity
        
        models distribution - absolutel cosine val has no meaning e.g. 0.5 isnt 'half right'
    - - distribution of values, absolute cosine si
    - - hybrid search
        
        keyword (BM25)
        
        matches literal tokens (stems)
        
        score = frequency appears in doc, weighted by how rare each term is. No motion of 'meaning'
        
        keyword wins: "What does error code E-4021 mean?"
        
        +vector search
        
        score = cloeness in embedding space
        
        'declined', 'rejected' 'wouldnt go through; will be scored similar
        
        vector search wins: "My card got declined" → doc says "the transaction was rejected by the issuing bank"
        
        Index/ANN/HNSW
        
        data structure that makes comparing against milllions fast
        
        keyword = precise on exact terms, brittle on wording; vector = robust to wording, blurry on exact terms
        
        RRF
        
        0-1 and 0-30+. RRG ignores 'score' and keeps rank position
      - reranker
        
        cross encoder?
      - chunking
        
        small coherent passes "orange-red breast" - useful at doc scale
      - HNSW - Hierarchical Navigable Small World. It's a data structure that makes vector search fast.
    - - tokens
      - latency