mirror of https://github.com/hpcaitech/ColossalAI
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
169 lines
7.6 KiB
169 lines
7.6 KiB
"""
|
|
Implement a memory class for storing conversation history
|
|
Support long term and short term memory
|
|
"""
|
|
from typing import Any, Dict, List
|
|
|
|
from colossalqa.chain.memory.summary import ConversationSummaryMemory
|
|
from colossalqa.chain.retrieval_qa.load_chain import load_qa_chain
|
|
from langchain.chains.combine_documents.base import BaseCombineDocumentsChain
|
|
from langchain.memory.chat_message_histories.in_memory import ChatMessageHistory
|
|
from langchain.schema import BaseChatMessageHistory
|
|
from langchain.schema.messages import BaseMessage
|
|
from langchain.schema.retriever import BaseRetriever
|
|
from pydantic import Field
|
|
|
|
|
|
class ConversationBufferWithSummary(ConversationSummaryMemory):
|
|
"""Memory class for storing information about entities."""
|
|
|
|
# Define dictionary to store information about entities.
|
|
# Store the most recent conversation history
|
|
buffered_history: BaseChatMessageHistory = Field(default_factory=ChatMessageHistory)
|
|
# Temp buffer
|
|
summarized_history_temp: BaseChatMessageHistory = Field(default_factory=ChatMessageHistory)
|
|
human_prefix: str = "Human"
|
|
ai_prefix: str = "Assistant"
|
|
buffer: str = "" # Formated conversation in str
|
|
existing_summary: str = "" # Summarization of stale converstion in str
|
|
# Define key to pass information about entities into prompt.
|
|
memory_key: str = "chat_history"
|
|
input_key: str = "question"
|
|
retriever: BaseRetriever = None
|
|
max_tokens: int = 2000
|
|
chain: BaseCombineDocumentsChain = None
|
|
input_chain_type_kwargs: List = {}
|
|
|
|
@property
|
|
def buffer(self) -> Any:
|
|
"""String buffer of memory."""
|
|
return self.buffer_as_messages if self.return_messages else self.buffer_as_str
|
|
|
|
@property
|
|
def buffer_as_str(self) -> str:
|
|
"""Exposes the buffer as a string in case return_messages is True."""
|
|
self.buffer = self.format_dialogue()
|
|
return self.buffer
|
|
|
|
@property
|
|
def buffer_as_messages(self) -> List[BaseMessage]:
|
|
"""Exposes the buffer as a list of messages in case return_messages is False."""
|
|
return self.buffered_history.messages
|
|
|
|
def clear(self):
|
|
"""Clear all the memory"""
|
|
self.buffered_history.clear()
|
|
self.summarized_history_temp.clear()
|
|
|
|
def initiate_document_retrieval_chain(
|
|
self, llm: Any, prompt_template: Any, retriever: Any, chain_type_kwargs: Dict[str, Any] = {}
|
|
) -> None:
|
|
"""
|
|
Since we need to calculate the length of the prompt, we need to initiate a retrieval chain
|
|
to calculate the length of the prompt.
|
|
Args:
|
|
llm: the language model for the retrieval chain (we won't actually return the output)
|
|
prompt_template: the prompt template for constructing the retrieval chain
|
|
retriever: the retriever for the retrieval chain
|
|
max_tokens: the max length of the prompt (not include the output)
|
|
chain_type_kwargs: the kwargs for the retrieval chain
|
|
memory_key: the key for the chat history
|
|
input_key: the key for the input query
|
|
"""
|
|
self.retriever = retriever
|
|
input_chain_type_kwargs = {k: v for k, v in chain_type_kwargs.items() if k not in [self.memory_key]}
|
|
self.input_chain_type_kwargs = input_chain_type_kwargs
|
|
self.chain = load_qa_chain(llm, chain_type="stuff", prompt=prompt_template, **self.input_chain_type_kwargs)
|
|
|
|
@property
|
|
def memory_variables(self) -> List[str]:
|
|
"""Define the variables we are providing to the prompt."""
|
|
return [self.memory_key]
|
|
|
|
def format_dialogue(self, lang: str = "en") -> str:
|
|
"""Format memory into two parts--- summarization of historical conversation and most recent conversation"""
|
|
if len(self.summarized_history_temp.messages) != 0:
|
|
for i in range(int(len(self.summarized_history_temp.messages) / 2)):
|
|
self.existing_summary = (
|
|
self.predict_new_summary(
|
|
self.summarized_history_temp.messages[i * 2 : i * 2 + 2], self.existing_summary, stop=["\n\n"]
|
|
)
|
|
.strip()
|
|
.split("\n")[0]
|
|
.strip()
|
|
)
|
|
for i in range(int(len(self.summarized_history_temp.messages) / 2)):
|
|
self.summarized_history_temp.messages.pop(0)
|
|
self.summarized_history_temp.messages.pop(0)
|
|
conversation_buffer = []
|
|
for t in self.buffered_history.messages:
|
|
if t.type == "human":
|
|
prefix = self.human_prefix
|
|
else:
|
|
prefix = self.ai_prefix
|
|
conversation_buffer.append(prefix + ": " + t.content)
|
|
conversation_buffer = "\n".join(conversation_buffer)
|
|
if len(self.existing_summary) > 0:
|
|
if lang == "en":
|
|
message = f"A summarization of historical conversation:\n{self.existing_summary}\nMost recent conversation:\n{conversation_buffer}"
|
|
elif lang == "zh":
|
|
message = f"历史对话概要:\n{self.existing_summary}\n最近的对话:\n{conversation_buffer}"
|
|
else:
|
|
raise ValueError("Unsupported language")
|
|
return message
|
|
else:
|
|
message = conversation_buffer
|
|
return message
|
|
|
|
def get_conversation_length(self):
|
|
"""Get the length of the formatted conversation"""
|
|
prompt = self.format_dialogue()
|
|
length = self.llm.get_num_tokens(prompt)
|
|
return length
|
|
|
|
def load_memory_variables(self, inputs: Dict[str, Any]) -> Dict[str, str]:
|
|
"""Load the memory variables.
|
|
Summarize oversize conversation to fit into the length constraint defined by max_tokene
|
|
Args:
|
|
inputs: the kwargs of the chain of your definition
|
|
Returns:
|
|
a dict that maps from memory key to the formated dialogue
|
|
the formated dialogue has the following format
|
|
if conversation is too long:
|
|
A summarization of historical conversation:
|
|
{summarization}
|
|
Most recent conversation:
|
|
Human: XXX
|
|
Assistant: XXX
|
|
...
|
|
otherwise
|
|
Human: XXX
|
|
Assistant: XXX
|
|
...
|
|
"""
|
|
# Calculate remain length
|
|
if "input_documents" in inputs:
|
|
# Run in a retrieval qa chain
|
|
docs = inputs["input_documents"]
|
|
else:
|
|
# For test
|
|
docs = self.retriever.get_relevant_documents(inputs[self.input_key])
|
|
inputs[self.memory_key] = ""
|
|
inputs = {k: v for k, v in inputs.items() if k in [self.chain.input_key, self.input_key, self.memory_key]}
|
|
prompt_length = self.chain.prompt_length(docs, **inputs)
|
|
remain = self.max_tokens - prompt_length
|
|
while self.get_conversation_length() > remain:
|
|
if len(self.buffered_history.messages) <= 2:
|
|
raise RuntimeError("Exceed max_tokens, trunk size of retrieved documents is too large")
|
|
temp = self.buffered_history.messages.pop(0)
|
|
self.summarized_history_temp.messages.append(temp)
|
|
temp = self.buffered_history.messages.pop(0)
|
|
self.summarized_history_temp.messages.append(temp)
|
|
return {self.memory_key: self.format_dialogue()}
|
|
|
|
def save_context(self, inputs: Dict[str, Any], outputs: Dict[str, str]) -> None:
|
|
"""Save context from this conversation to buffer."""
|
|
input_str, output_str = self._get_input_output(inputs, outputs)
|
|
self.buffered_history.add_user_message(input_str.strip())
|
|
self.buffered_history.add_ai_message(output_str.strip())
|