Skip to content

Comment on Contextual Retrieval

Comments

If your knowledge base is smaller than 200,000 tokens (about 500 pages of material)

I would prefer that anthropic just release their tokeniser so we don't have to make guesses.

shouldn't this be possible to reverse-engineer since they stream the responses token-by-token?

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.