Types

Manages a stateful conversation between a user and a language model.

The ConversationChain combines three key components:

1. memory: Stores and retrieves conversation history
2. llm: The language model that generates responses
3. prompt: Template for formatting the conversation for the LLM

When invoked with a user message, the ConversationChain:

• Adds the user message to memory
• Retrieves the updated conversation history
• Formats the conversation for the LLM using the prompt template
• Gets a response from the LLM
• Stores the AI response in memory
• Returns the AI response

Example:

import Data.Text (Text)
import qualified Data.Text as T
import Langchain.LLM.OpenAI (OpenAI(..))
import Langchain.Memory.ConversationBufferMemory (ConversationBufferMemory(..))
import Langchain.PromptTemplate (PromptTemplate(..), createPromptTemplate)
import Langchain.Runnable.ConversationChain (ConversationChain(..))

main :: IO ()
main = do
  -- Create memory component
  let memory = ConversationBufferMemory
        { messages = []
        , returnMessages = True
        }

  -- Create LLM
  let llm = OpenAI
        { model = "gpt-4"
        , temperature = 0.7
        }

  -- Create prompt template
  promptTemplate <- createPromptTemplate
    "You are a helpful assistant. {history}\nHuman: {input}\nAI:"
    ["history", "input"]

  -- Create conversation chain
  let conversation = ConversationChain
        { memory = memory
        , llm = llm
        , prompt = promptTemplate
        }

  -- Start conversation
  response1 <- invoke conversation "Hello, who are you?"
  case response1 of
    Left err -> putStrLn $ "Error: " ++ T.unpack err
    Right answer -> do
      putStrLn $ "AI: " ++ T.unpack answer

      -- Continue conversation with context
      response2 <- invoke conversation "What can you help me with?"
      case response2 of
        Left err -> putStrLn $ "Error: " ++ T.unpack err
        Right answer2 -> putStrLn $ "AI: " ++ T.unpack answer2

You can customize the behavior by using different memory implementations:

• ConversationBufferMemory - Stores the full conversation history
• ConversationBufferWindowMemory - Keeps only the most recent N exchanges
• ConversationSummaryMemory - Summarizes older conversations to save tokens
• ConversationEntityMemory - Tracks entities mentioned in the conversation

The prompt template can be customized to give the LLM specific instructions, persona characteristics, or to format the conversation history in different ways.