1

Hello, novice here. Does gpt-3.5-turbo model has built-in long-term memory capabilities? Or do I need to glue the previous msg sent like I would with davinci? Thank you for your answer.

1 Like
3

Thank you very much for your answer Raul. What I mean is not to tailor to specific users, but to have long term memory, remember previous msgs.

5

Raul, do you use which model? davinci or gpt turbo?

7

These docs will probably be helpful: OpenAI API.

You send the whole conversation history with e.g.

# Note: you need to be using OpenAI Python v0.27.0 for the code below to work
import openai

openai.ChatCompletion.create(
  model="gpt-3.5-turbo",
  messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Who won the world series in 2020?"},
        {"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."},
        {"role": "user", "content": "Where was it played?"}
    ]
)
1 Like
8

@dliden thanks for the referece. This raises the problem of how to maintain long conversations due to limitations of the context window and pricing.

1 Like
9

Those are open questions a lot of people are working on. You might want to check out e.gt. GitHub - zilliztech/GPTCache: GPTCache is a library for creating semantic cache to store responses from LLM queries. and Memory — 🦜🔗 LangChain 0.0.139.

2 Likes
10

thanks dliden, didn’t know about gptcache!

2 Likes
11

We built simple & semantic cache recently for our tool - ⭐ Reducing LLM Costs & Latency with Semantic Cache

It’s very easy to use it in production - just requires changing the base URL to Portkey (our platform) and passing a cache header in your request.

We are seeing a consistent 20% cache hit rate for Q&A and RAG use cases with 99% accuracy.