Ikibazo gihuza ibibazo · Hagati Urwego

Ikibazo cyo Gusubiramo

Gerageza ibizamini bisubirwamo: gutegura inyandiko zijyanye nibice, gushiramo imiterere hamwe na BM25, guhinduranya, hamwe nigiciro cyo guhita.

Inzira zijyanye no kuyoboraGusubiramo
Ikibazo 1 Bya 8

What problem does contextual retrieval solve?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. What problem does contextual retrieval solve?
    • Chunks lose information like company or time period when split from their document
    • Embeddings are too expensive to store
    • Language models cannot read tables
    • BM25 cannot handle long queries
  2. How is the context for each chunk generated?
    • By copying the document title only
    • A language model reads the whole document and the chunk and writes a short situating explanation
    • By a human editor for every chunk
    • By averaging neighboring embeddings
  3. Which two indexes receive the contextualized chunk text in Anthropic's approach?
    • A graph database and a cache
    • A SQL table and a spreadsheet
    • An embedding index and a BM25 keyword index
    • Two separate vector indexes with different models
  4. According to Anthropic's reported experiments, roughly how much did contextual embeddings plus contextual BM25 reduce top-20 retrieval failures?
    • About 5 percent
    • About 20 percent
    • About 90 percent
    • About 49 percent
  5. Why does prompt caching make contextual retrieval cheaper?
    • It caches the document once and reuses it for every chunk's context call
    • It skips the embedding step
    • It removes the need for BM25
    • It shortens each chunk
  6. Why is keeping BM25 in the pipeline especially useful with added context?
    • BM25 understands synonyms better than embeddings
    • Contexts often add exact identifiers like names, dates or codes that keyword search matches precisely
    • BM25 requires no index
    • BM25 replaces reranking
  7. When might contextual retrieval be unnecessary?
    • When the documents are in English
    • When using a reranker
    • When the knowledge base is small enough to fit in the model's context window
    • When chunks are longer than 50 tokens
  8. Where does the main extra cost of contextual retrieval occur?
    • At query time for every user question
    • In the reranker only
    • In the user's browser
    • At indexing time, with one model call per chunk