Gids-gekoppelde quiz · Middelmatig Niveau
Lange context versus RAG-quiz
Vergelijk lange context en RAG op kosten, latentie, nauwkeurigheid en versheid, inclusief het lost-in-the-middle-effect en hybride ontwerpen.
Vraag 1 van 8
Waarom kost de lange-contextbenadering doorgaans meer per zoekopdracht dan RAG?
Bekijk elke vraag8 vragen
Antwoorden en uitleg verschijnen terwijl je de quiz doet.
- Waarom kost de lange-contextbenadering doorgaans meer per zoekopdracht dan RAG?
- Providers charge per input token, and long context sends far more tokens
- Long-context models require a vector database
- RAG is always free
- Long context requires fine-tuning for each question
- Wat heeft het Lost in the Middle-onderzoek uit 2023 gevonden?
- Models ignore the first sentence of every prompt
- Retrieval always beats long context
- Performance dropped when relevant information was placed in the middle of long inputs
- Context windows cannot exceed 4,000 tokens
- Wat is een belangrijke beperking van naald-in-een-hooiberg-tests?
- Ze werken alleen met afbeeldingen
- They measure simple lookup, not reasoning across many scattered facts
- They cannot be run on long inputs
- They measure cost, not accuracy
- Welke situatie geeft het duidelijkst de voorkeur aan RAG boven een lange context?
- A single short policy document queried a few times
- Summarizing the themes of one novel
- Comparing two chapters of one report
- A large, frequently changing corpus with per-user access permissions
- Hoe moet de prompt worden gerangschikt bij gebruik van promptcaching met een lang corpus?
- Put the question first and the corpus last
- Shuffle the corpus on each request
- Put the stable corpus first and the variable question last
- Split the corpus across separate requests
- Why does time to first token increase with context length?
- The model must process every input token before it starts answering
- Longer prompts are sent over slower networks
- The model retrieves documents from the internet
- Output tokens are generated in parallel
- Which task is long context particularly well suited to?
- Searching millions of support articles
- Serving users with different clearance levels
- Answering from data that changes every hour
- Summarizing themes across an entire document
- What does a common hybrid of long context and RAG look like?
- Using RAG only for images and long context only for text
- Retrieving relevant documents generously, then including whole documents in a long window
- Fine-tuning the model on the retrieved chunks
- Alternating approaches on odd and even days