RecallM attempts to bring us one step closer to achieving Artficial General Intelligence (AGI) by providing Large Language Models (LLMs) with an adaptable and updatable long-term memory.
RecallM is intended to work in the same manner as a typical chatbot, while retaining any new information provided to it. Users are able to update the knowledge of the system in natural language.
RecallM offers a complete solution encapsulated into a single python object using the RecallM architecture with GPT via LangChain. However, we also offer a LangChain document retriever implementation of RecallM to allow for complete modularity and flexibility with other LangChain applications.
Click here to see complete details about the RecallM architecture.
- Clone the repository and submodules:
> git clone git@github.com:cisco-open/DeepVision.git
> cd DeepVision/recallm
> git submodule update --init --recursive
- Create the api keys json file by executing the following command inside the recallm folder. Replace <API_KEY> with your OpenAI api key:
> echo {"open_ai":"<API_KEY>"} > api_keys.json
- Ensure that you have Docker installed, then create the Docker container:
> docker compose up -d
- Pip install required packages: (Ideally in a new environment)
> pip install -r requirements.txt
Ensure that the docker container is running, then use the RecallM system as follows:
from recall import RecallM
# Initialize the RecallM instance
recallm = RecallM()
# Add knowledge from a string
recallm.update(knowledge="Brandon loves coffee")
# Add knowledge from a web page (Warning: this might miss some text that is dynamically loaded)
recallm.update_from_url(url='https://paperswithcode.com/paper/recallm-an-architecture-for-temporal-context')
# Add knowledge from a text file
recallm.update_from_file(file='./datasets/other/state_of_the_union.txt')
question = "What do you know about Brandon?"
answer = recallm.question(question)
print(answer)
# Always close the Docker database connection when finished
recallm.close()Alternatively, to quickly start using RecallM you can launch the interactive terminal interface using:
> python recall_terminal.py
RecallM also implements the BaseRetriever class from LangChain so that it can be implemented in other LangChain applications:
from langchain.chains import RetrievalQAWithSourcesChain
from langchain.chat_models import ChatOpenAI
from recall import RecallM
OPEN_AI_API_KEY = '....'
recallm = RecallM(openai_key=OPEN_AI_API_KEY)
recallm.update(knowledge="Brandon loves coffee")
chat = ChatOpenAI(temperature=0,
openai_api_key=OPEN_AI_API_KEY,
model_name="gpt-3.5-turbo")
chain = RetrievalQAWithSourcesChain.from_chain_type(
chat,
chain_type="stuff",
retriever=recallm.as_retriever()
)
question = "What do you know about Brandon?"
answer = chain({"question": question}, return_only_outputs=True)['answer']
print(answer)
# Always close the Docker database connection when finished
recallm.close()recallm.reset_knowledge()Please note that by simply running Neo4J desktop instead of the docker container, you can visualize the knowledge graph being created with RecallM.
Please cite our work using:
@misc{kynoch2023recallm,
title={RecallM: An Architecture for Temporal Context Understanding and Question Answering},
author={Brandon Kynoch and Hugo Latapie},
year={2023},
eprint={2307.02738},
archivePrefix={arXiv},
primaryClass={cs.AI}
}