07. Vector Storage Search Memory (VectorStoreRetrieverMemory)

VectorStoreRetrieverMemory

VectorStoreRetrieverMemory Saves memory in the vector store and queries the most'noticeable' top K documents each time it is called.

This is the order of conversation Not explicitly tracking It is different from most other memory classes in.

Copy

# API KEY A configuration file for managing environment variables
from dotenv import load_dotenv

# API KEY load information
load_dotenv()

Copy

True

Copy

First, initialize the vector store.

Copy

import faiss
from langchain_openai import OpenAIEmbeddings
from langchain.docstore import InMemoryDocstore
from langchain.vectorstores import FAISS


# Define an embedding model.
embeddings_model = OpenAIEmbeddings()

# Vector Store Initializes.
embedding_size = 1536
index = faiss.IndexFlatL2(embedding_size)
vectorstore = FAISS(embeddings_model, index, InMemoryDocstore({}), {})

In real use k Set to a higher value, k=1 Use to display as follows:

Copy

from langchain.memory import VectorStoreRetrieverMemory

# To show that vector lookups still return semantically relevant information.
retriever = vectorstore.as_retriever(search_kwargs={"k": 1})
memory = VectorStoreRetrieverMemory(retriever=retriever)

# Save random conversations.
memory.save_context(
    inputs={
        "human": "Hello, thank you for coming to the interview today. Please introduce yourself."
    },
    outputs={
        "ai": "Hello. I am a new developer majoring in computer science. In college, I mainly used Java and Python, and recently I participated in a web development project and gained experience developing services for real users."
    },
)
memory.save_context(
    inputs={"human": "What role did you play in the project?"},
    outputs={
        "ai": "My role was as a backend developer. I was in charge of processing user data and developing server logic, and I was in charge of implementing RESTful APIs to communicate with the frontend. I also participated in database design."
    },
)
memory.save_context(
    inputs={
        "human": "If you have ever encountered difficulties on a team project, how did you resolve them?"
    },
    outputs={
        "ai": "There were some difficulties due to communication issues at the beginning of the project. To solve this, our team had regular meetings and shared each other's progress. Also, when problems occurred, we actively shared opinions and tried to find reasonable solutions."
    },
)
memory.save_context(
    inputs={"human": "What do you think are your strengths as a developer?"},
    outputs={
        "ai": "My strengths are my ability to learn quickly and problem-solve. I can quickly learn new technologies or tools, and I can come up with creative solutions when faced with complex problems. I also value teamwork and consider it important to collaborate with my colleagues."
    },
)

Returns the most relevant conversation in the Vector Store lot 1 (because it is k=1) when asked the following questions:

• Question: "What is your interviewer major?"

Copy

# Extract the single most relevant conversation by asking questions in memory.
print(memory.load_memory_variables({"prompt": "면접자 전공은 무엇인가요?"})["history"])

Copy

human: Hello, thank you for attending the interview today. Please introduce yourself. 
ai: Hello. I am a new developer majoring in computer science. The university mainly used Java and Python, and recently I participated in a web development project and had the experience of developing services for real users. 

This time, we extract the most relevant 1 conversation through other questions.

• Question: "What role did the interviewer play in the project?"

Copy

print(
    memory.load_memory_variables(
        {"human": "What role does the interviewer play in the project?"}
    )["history"]
)

Copy

human: What role did you play in the project? 
ai: My role was backend developer. He was responsible for processing user data and developing server logic, and implemented the RESTful API to communicate with the front end. He also participated in database design.