how.wtf

Custom memory in Langchain

· Thomas Taylor

Implementing custom memory in Langchain is dead simple using the ChatMessageHistory class.

How to implement custom memory in Langchain (including LCEL)

One of the easiest methods for storing and retrieving messages with Langchain is using the ChatMessageHistory class that is provided from the langchain.memory module.

It’s simple to get started with:

 1from langchain.memory import ChatMessageHistory
 2
 3history = ChatMessageHistory()
 4
 5history.add_user_message("Hello!")
 6
 7history.add_ai_message("Yo!")
 8
 9print(history.messages)
10
11# [HumanMessage(content='Hello!'), AIMessage(content='Yo!')]

The interface for ChatMessageHistory is: add_user_message and add_ai_message.

For the sake of this example, I’m using a .json file with preloaded messages. A database, file, etc. may be used instead.

 1{
 2    "messages": [
 3        {
 4            "sender": "human",
 5            "body": "Hello!"
 6        },
 7        {
 8            "sender": "ai",
 9            "body": "Yo!"
10        }
11    ]
12}
 1import json
 2
 3from langchain.memory import ChatMessageHistory
 4
 5with open("messages.json") as json_file:
 6    content = json.load(json_file)
 7
 8messages = content["messages"]
 9
10chat_history = ChatMessageHistory()
11for m in messages:
12    if m["sender"] == "human":
13        chat_history.add_user_message(m["body"])
14    elif m["sender"] == "ai":
15        chat_history.add_ai_message(m["body"])
16
17print(chat_history.messages)
18# [HumanMessage(content='Hello!'), AIMessage(content='Yo!')]

The chat_history may be used for instantiating other types of memory!

Here is an example from the langchain documentation using the ChatMessageHistory with Langchain Expression Language (LCEL):

 1import json
 2
 3from operator import itemgetter
 4
 5from langchain.chat_models import ChatOpenAI
 6from langchain.memory import ConversationBufferMemory
 7from langchain.schema.runnable import RunnablePassthrough, RunnableLambda
 8from langchain.prompts import ChatPromptTemplate, MessagesPlaceholder
 9
10with open("messages.json") as json_file:
11    content = json.load(json_file)
12
13messages = content["messages"]
14
15chat_history = ChatMessageHistory()
16for m in messages:
17    if m["sender"] == "human":
18        chat_history.add_user_message(m["body"])
19    elif m["sender"] == "ai":
20        chat_history.add_ai_message(m["body"])
21
22model = ChatOpenAI()
23prompt = ChatPromptTemplate.from_messages(
24    [
25        ("system", "You are a helpful chatbot"),
26        MessagesPlaceholder(variable_name="history"),
27        ("human", "{input}"),
28    ]
29)
30
31memory = ConversationBufferMemory(return_messages=True, chat_memory=chat_history)
32
33chain = (
34    RunnablePassthrough.assign(
35        history=RunnableLambda(memory.load_memory_variables) | itemgetter("history")
36    )
37    | prompt
38    | model
39)
40
41inputs = {"input": "hi im bob"}
42response = chain.invoke(inputs)

#Python   #Generative-Ai  

Reply to this post by email ↪