search

simonw / llm-claude-3

LLM plugin for interacting with the Claude 3 family of models
Apache License 2.0
223 stars 21 forks source link

Does `llm -c` re-send the whole prompt? #5

Closed almson closed 4 months ago

almson commented 5 months ago

I am experimenting with Opus, and my input token usage seems very high considering that I've only sent the main data once and then used "llm -c". Does this plugin keep a conversation going, or does it re-send the whole prompt each time?

pointlessone commented 4 months ago

It does. So it does for each message in chat. This is how all the LLMs work: they need the whole conversation in the context in order to see it, they don't have any memory.

simonw commented 4 months ago

Yes, it re-sends the whole prompt - as do most of the other LLM plugins.

almson commented 4 months ago

It does. So it does for each message in chat. This is how all the LLMs work: they need the whole conversation in the context in order to see it, they don't have any memory.

Yeah... that doesn't matter.

It's about how the API is built and how billing is calculated. ChatGTP has a session-based API, but it seems Claude doesn't.

That's very unfortunate, especially when I'm trying to work with eg a 100k token codebase. The expense becomes unmanageable. I'll have to stick to the web UI for now.

almson commented 3 months ago

I'm sorry, I had a wrong idea about OpenAI and Anthropic APIs. Unfortunately, though, it means that if I want to upload a codebase, it will be much cheaper to use Claude's web UI which does not charge me to process the whole codebase on each request.