store

Store

VectorStore

The VectorStore is storage of embeddings. It should follow API design below to adapt operations in chatbot:

VectorStore(table_name: str, embedding_func: Embeddings)

Parameters:

• table_name (str): table name in vector database
• embedding_func (Embeddings): embedding method to convert query or documents to vectors

Methods:

• insert: data insert, given a list of documents, returns how many data entities inserted
• search: semantic search, given a query in string, returns a list of useful documents

By default, it uses Milvus in LangChain. You can modify config.py to configure it. The default module also works with Zilliz Cloud by setting configurations for the vector store below:

# Vector db configs
VECTORDB_CONFIG = {
    'connection_args': {
        'uri': os.getenv('MILVUS_URI', 'your_endpoint'),
        'user': os.getenv('MILVUS_USER', 'user_name'),
        'password': os.getenv('MILVUS_PASSWORD', 'password_goes_here'),
        'secure': True
        },
    'top_k': 10,
    'index_params': {
        'metric_type': 'IP',
        'index_type': 'AUTOINDEX',
        'params': {}
        }
}

ScalarStore (Optional)

The ScalarStore is storage of scalar data, which allows information retrieval other than semantic search, such as keyword match. It should follow API design below to adapt operations in chatbot:

ScalarStore(index_name: str, client: CLIENT)

Parameters:

• index_name (str): table name in scalar database
• client: method to connect database

Methods:

• insert: data insert, given a list of documents, returns how many data entities inserted
• search: scalar search, given a query in string, returns a list of useful documents

To enable scalar store, you need to set USE_SCALAR=True in config.py. By default, it uses ElasticSearch BM25 in LangChain. You can modify config.py to configure connection args.

MemoryStore

The MemoryStore records chat history in database. It should follow API design below to adapt operations in chatbot:

MemoryStore(table_name: str, session_id: str)

Parameters:

• table_name (str): table name in database
• session_id (str): identifier for sessions, allowing for different sessions of conversation

Attributes:

• memory (BaseMemory): a LangChain base memory to adapt agent

Methods:

• add_history: insert chat history to database, given a list of dictionaries with keys of 'question' and 'answer', [{'question': 'xxx', 'answer': 'xxx'}]
• get_history: return chat history in a list of tuples, [('this is question', 'this is answer')]

By default, it uses PostgresChatMessageHistory and ConversationBufferMemory in LangChain to build memory. You can modify config.py to configure it.