-
Notifications
You must be signed in to change notification settings - Fork 164
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
added memory documentation #102
base: main
Are you sure you want to change the base?
added memory documentation #102
Conversation
WalkthroughThe update introduces a comprehensive framework for memory management in AI conversational systems. It outlines various memory classes, each tailored to handle different aspects of memory storage and retrieval, ensuring efficient and scalable conversation handling. The document also discusses how these memory components interact with other system elements and hints at possible enhancements. Changes
Thank you for using CodeRabbit. We offer it for free to the OSS community and would appreciate your support in helping us grow. If you find it useful, would you consider giving us a shout-out on your favorite social media? TipsChatThere are 3 ways to chat with CodeRabbit's AI:
Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments. CodeRabbit Commands (invoked as PR comments)
Additionally, you can add CodeRabbit Configration File (
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Review Status
Actionable comments generated: 9
Configuration used: CodeRabbit UI
Files selected for processing (1)
- docs/ai/03_memory.md (1 hunks)
Additional comments: 1
docs/ai/03_memory.md (1)
- 2-4: The function
extract_text
is described as a standalone function, but it's not clear from the documentation where this function is defined or how it's related to the memory classes. Clarify the context in whichextract_text
is used within the memory management system.
class SimpleMemory(BaseMemory, BaseModel): - Defines the SimpleMemory | ||
|
||
class that inherits from BaseMemory. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The SimpleMemory
class is introduced without a clear description of its purpose or how it differs from the other memory classes. Provide a brief description of the class's role in memory management.
def add\_memory(self, prompt: str, llm\_response: Any) -> None: - Adds a conversation memory to the store. | ||
|
||
Parameters: | ||
|
||
prompt (str): The user's prompt or input. | ||
|
||
llm\_response (Any): The AI's response to the prompt. | ||
|
||
Functionality: | ||
|
||
This method checks if the given (prompt, llm\_response) pair already exists in memory. | ||
|
||
If not, it appends a dictionary containing the prompt and response to the list of self.messages. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The add_memory
method in SimpleMemory
is well-documented, but it's important to specify what type of store is being used. Is it an in-memory list, a database, or a file system? This information is crucial for understanding the method's implementation and potential limitations.
def get\_memory(self, \*\*kwargs) -> str: - Retrieves all memories from the store. | ||
|
||
Parameters: | ||
|
||
kwargs (dict): Additional keyword arguments (not used in this method). | ||
|
||
Functionality: | ||
|
||
It constructs a formatted string containing all conversations stored in self.messages. | ||
|
||
For each conversation, it appends the user's prompt and AI's response to the formatted string. | ||
|
||
Returns the formatted string containing all memories. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The get_memory
method's description does not specify the format of the returned string. For better clarity, include an example of the formatted string or describe the formatting in more detail.
def remove\_memory(self, prompt: str) -> None: - Removes a specific memory from the store. | ||
|
||
Parameters: | ||
|
||
prompt (str): The user's prompt to be removed. | ||
|
||
Functionality: | ||
|
||
Iterates through each conversation in self.messages. If a conversation's prompt matches the provided prompt, it is removed from the list. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The remove_memory
method description should clarify what happens if the prompt is not found in the memory. Does it fail silently, or is an error raised? This information is important for error handling and expected behavior.
class SummaryMemory(BaseMemory, BaseModel): - Defines the SummaryMemory class that inherits from BaseMemory. | ||
|
||
Attributes: | ||
|
||
current\_summary (str): Keeps track of the current conversation summary. | ||
|
||
messages\_in\_summary (List[Dict[BaseMessage, Any]]): Stores a list of messages that have been included in the summary. | ||
|
||
Methods | ||
|
||
Method add\_memory: | ||
|
||
This method adds a new conversation memory to the store, avoiding duplicates. | ||
|
||
Method get\_memory: | ||
|
||
It creates an instance of the llms.OpenAI class using the compiler module. It filters out messages that are already included in the current summary. For each new message, it constructs a string with the user prompt and AI response and updates messages\_in\_summary. | ||
|
||
It creates a summarizer using the compiler module with the provided template, llm instance, and new lines. The summary is updated with the extracted text from the summarized memory. | ||
|
||
The final summarized memory is returned. | ||
|
||
Method remove\_memory: | ||
|
||
This method removes a specified memory (user prompt) from the store. It removes the memory from messages, updates current\_summary based on the remaining messages, and clears messages\_in\_summary. | ||
|
||
Method clear: | ||
|
||
Clears all stored memories, the list of messages in the summary, and resets the current summary. | ||
|
||
The SummaryMemory class builds upon the concept of storing and managing conversation memories. It adds the capability to create and update summaries of conversations using a summarization process. The class utilizes the compiler module to interact with OpenAI's language model and performs summarization based on a predefined template. It keeps track of which messages are included in the summary to prevent duplication. The SummaryMemory class seems to be designed for a more advanced conversational AI system that can generate summaries of interactions. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The SummaryMemory
class description mentions the use of the compiler module
and interaction with OpenAI's language model
. Ensure that the documentation includes links to the relevant sections or external documentation for these components, as they are critical to understanding how SummaryMemory
operates.
class BufferSummaryMemory(BaseMemory, BaseModel): - Defines the BufferSummaryMemory class that inherits from BaseMemory. | ||
|
||
Attributes: | ||
|
||
current\_summary (str): Tracks the current conversation summary. | ||
|
||
current\_buffer (str): Stores recent messages in a buffer. | ||
|
||
messages\_in\_summary (List[Dict[BaseMessage, Any]]): Stores messages included in the summary. | ||
|
||
messages\_in\_buffer (List[Dict[BaseMessage, Any]]): Stores messages included in the buffer. | ||
|
||
Methods | ||
|
||
Method add\_memory: | ||
|
||
Adds a new conversation memory to the store while avoiding duplicates. | ||
|
||
Method get\_memory: | ||
|
||
It deals with both the summary and the buffer. Creates an instance of the llms.OpenAI class using the compiler module. Retrieves the memory\_threshold from kwargs or defaults to 1. Separates recent messages into the buffer and older ones into the summary based on the threshold. Constructs the buffer as a string. Generates a summary only for the new additions to the summary and updates the current\_summary. Returns the combined summary and buffer as the summarized memory. | ||
|
||
Method remove\_memory: | ||
|
||
Removes a specific memory (user prompt) from the store and updates related variables. | ||
|
||
Recalculates the buffer and summary based on the remaining messages and threshold, and updates the current\_buffer and current\_summary accordingly. | ||
|
||
Method clear: | ||
|
||
Clears all stored memories, the lists of messages in the summary and buffer, and resets the current summary and buffer. | ||
|
||
The BufferSummaryMemory class extends the functionality of SummaryMemory by introducing a buffer to store recent messages separately from the summary. This design allows the class to manage recent conversations in a buffer and maintain a more concise summary of older interactions. It utilizes a threshold to decide which messages are included in the buffer and which in the summary. The class's methods manage both the buffer and summary, offering a more sophisticated approach to memory management for conversational AI systems that require both a concise summary and recent context. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The BufferSummaryMemory
class extends SummaryMemory
, but it's not clear how the buffer mechanism is implemented or what the threshold for separating recent messages is. Provide more details on the buffer mechanism and how the threshold value is determined or can be configured.
|
||
Method add\_memory: | ||
|
||
def add\_memory(self, program\_name: str, session\_id: str, prompt: str, llm\_response: str) -> None: - Adds a memory entry to the store for a specific session of an application. | ||
|
||
Parameters: | ||
|
||
program\_name (str): Name of the application. | ||
|
||
session\_id (str): Identifier for the specific session. | ||
|
||
prompt (str): User's input prompt. | ||
|
||
llm\_response (str): AI's response to the prompt. | ||
|
||
Functionality: | ||
|
||
Creates a directory for the specific program if it doesn't exist. Writes the given prompt and response as a JSON entry to the session file. | ||
|
||
Method get\_memory: | ||
|
||
def get\_memory(self, program\_name: str, session\_id: str, max\_len: int = 2040 \* 1024, limit: int = 1000): - Retrieves memories from a specific session of an application. | ||
|
||
Parameters: | ||
|
||
program\_name (str): Name of the application. | ||
|
||
session\_id (str): Identifier for the specific session. | ||
|
||
max\_len (int): Maximum length of data to read (default is 2040 KB). | ||
|
||
limit (int): Maximum number of memories to retrieve (default is 1000). | ||
|
||
Functionality: | ||
|
||
Reads the last max\_len bytes of the session file. Extracts the last limit JSON entries as memories. Returns the retrieved memories. | ||
|
||
Method remove\_memory: | ||
|
||
def remove\_memory(self, program\_name: str, session\_id: str, prompt: str = ''): - Removes memories from a specific session of an application. | ||
|
||
Parameters: | ||
|
||
program\_name (str): Name of the application. | ||
|
||
session\_id (str): Identifier for the specific session. | ||
|
||
prompt (str): If specified, removes only memories with the given prompt. | ||
|
||
Functionality: | ||
|
||
If prompt is not specified, it removes the entire session file. If prompt is specified, it removes only the memory entries with the specified prompt. | ||
|
||
Method clear: | ||
|
||
def clear(self, program\_name: str = ''): - Clears all memories of a specific program or all programs. | ||
|
||
Parameters: | ||
|
||
program\_name (str): Name of the application to clear memories for (optional). | ||
|
||
Functionality: | ||
|
||
If program\_name is specified, it clears all memories for that program. | ||
|
||
If program\_name is not specified, it iterates through all programs and clears their memories. | ||
|
||
Helper functions | ||
|
||
def read(self, program\_name: str, session\_id: str, n: int, max\_len: int) -> List[Dict]: - Reads and returns specific memories from a session of an application. | ||
|
||
Parameters: | ||
|
||
program\_name (str): Name of the application. | ||
|
||
session\_id (str): Identifier for the specific session. | ||
|
||
n (int): Number of memories to retrieve. | ||
|
||
max\_len (int): Maximum length of data to read. | ||
|
||
Functionality: | ||
|
||
Reads the last max\_len bytes of the session file. Extracts the last n JSON entries as memories. Returns the extracted memories. | ||
|
||
def sessions(self, program\_name: str) -> List[Dict]: - Retrieves a list of sessions for a specific program. | ||
|
||
Parameters: | ||
|
||
program\_name (str): Name of the application. | ||
|
||
Functionality: | ||
|
||
Searches for session files in the program's directory and retrieves information about each session, including the last modified time and latest memory entry. | ||
|
||
Method \_memories\_dirs: | ||
|
||
A private method that creates the directory for storing memory data if it doesn't exist and returns the directory path. | ||
|
||
The InFileMemory class provides a memory management system that saves conversation data in individual session files on the local filesystem. It's designed to persistently store conversations for different applications and sessions. It offers methods to add memories, retrieve memories, remove memories, clear memories, read specific memories, and retrieve session information. This class could be useful for maintaining a history of interactions in a more permanent storage location. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The InFileMemory
class documentation should include information about the file format used for storing sessions and how concurrency is handled. Are there any locking mechanisms or transactional guarantees when multiple instances try to write to the same session file?
class ReadOnlyMemory(BaseMemory): - Defines the ReadOnlyMemory class that inherits from BaseMemory. | ||
|
||
Attributes: | ||
|
||
memory: BaseMemory - Holds the reference to the original memory instance. | ||
|
||
Constructor: | ||
|
||
def \_\_init\_\_(self, memory: BaseMemory): - Initializes the ReadOnlyMemory instance with a provided memory instance. | ||
|
||
Parameters: | ||
|
||
memory (BaseMemory): The memory instance that this read-only memory will wrap. | ||
|
||
Methods | ||
|
||
Method add\_memory: | ||
|
||
def add\_memory(self, prompt: str, llm\_response: Any) -> None: - Prevents adding memory in a read-only context. | ||
|
||
Functionality: | ||
|
||
This method is implemented with an empty body, preventing any addition to the underlying memory instance. | ||
|
||
Method get\_memory: | ||
|
||
def get\_memory(self, \*\*kwargs) -> Any: - Retrieves entire memory from the store. | ||
|
||
Parameters: | ||
|
||
kwargs (dict): Additional keyword arguments that may be passed to the underlying memory's get\_memory method. | ||
|
||
Functionality: | ||
|
||
Calls the get\_memory method of the underlying memory instance and returns its result. | ||
|
||
Method remove\_memory: | ||
|
||
def remove\_memory(self, prompt: str) -> None: - Prevents memory removal in a read-only context. | ||
|
||
Functionality: | ||
|
||
This method is implemented with an empty body, preventing any removal from the underlying memory instance. | ||
|
||
Method clear: | ||
|
||
def clear(self) -> None: - Prevents clearing the memory in a read-only context. | ||
|
||
Functionality: | ||
|
||
This method is implemented with an empty body, preventing any clearing of the underlying memory instance. | ||
|
||
Property memory\_keys: | ||
|
||
@property def memory\_keys(self) -> List[str]: - Returns a list of prompts for all memories in the underlying memory. | ||
|
||
Functionality: | ||
|
||
Iterates through each conversation in the underlying memory and retrieves their prompts, creating a list of prompts. | ||
|
||
The ReadOnlyMemory class is a memory wrapper that provides a read-only view of an existing memory instance. It prevents modifications to the underlying memory while allowing data retrieval operations like getting all memories and extracting memory keys (prompts). This class is useful in scenarios where you want to ensure that memory data remains unmodified during certain operations or contexts. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The ReadOnlyMemory
class is described as a wrapper that prevents modifications. However, it would be beneficial to explain the use cases for such a class. When would one prefer to use ReadOnlyMemory
over the other memory classes?
Implementation with Compiler | ||
|
||
Compiler \_\_init\_\_.py | ||
|
||
Initializing a compiler takes an optional parameter ‘memory’, which should be a instance of one of the memory classes explained above. If memory is passed, the compiler passes it to the Program class. Additionally, if the user doesn’t prefer the conversation history to be displayed in the system prompt (in case the user doesn’t want a system prompt or wants to pass it with instructions, etc.), they can add a ConversationHistory variable to the prompt template. | ||
|
||
\_program.py | ||
|
||
1. Firstly, after all the variables are initialized, the class init function checks if memory has been passed. If yes, it further checks if the prompt template has a ConversationHistory variable, and if there isn’t, the ‘add\_variable’ function, declared in the \_program.py file itself, is called to add it to the prompt template at the end of the system prompt. | ||
|
||
1. Next, the value for ConversationHistory is added to the keyword arguments, (kwargs), by getting it from the Memory.get\_memory() function explained above. | ||
|
||
1. The memory variable is then passed to all Program class instances created for further use. | ||
|
||
1. After the response has been completely generated, the program again checks if memory exists. If yes, then it calls the extract\_text() function, declared the \_program.py file itself to get the user prompt and llm response key-value pair, which is then passed to Memory.add\_memory() to save it into the memory. | ||
|
||
Implementation with Agent | ||
|
||
The usage is same for agent as for the compiler. Simply create a memory instance and pass into the agent constructor. The agent \_\_init\_\_() function passes it into the compiler when the object is created, and the implementation continues as explained above. | ||
|
||
Future implementation | ||
|
||
Here are some suggestions on what could be improved on memory in the future. | ||
|
||
1. Manual saving of memory: Through a parameter, the user can disable the adding of prompt and llm responses by the Program instance, as explained in implementation with compiler, step 4. Instead, the user can call the memory.add\_memory() themselves if they wish to add only certain variables. This can be called either after the compiler/agent is run, or in the prompt | ||
|
||
template as well. | ||
|
||
1. Fixing infile memory: Unfortunately, while the infile memory has been created and works, it has not been integrated into the compiler, but it only requires the parameters to be taken from the user and to be passed into the add\_memory function, similar to how the memory\_threshold variable has been passed for BufferSummaryMemory. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The section on implementation with the Compiler and Agent is quite dense and could benefit from code snippets or pseudo-code to illustrate the described processes. Additionally, the future implementation suggestions should be clearly separated into their own subsection for better readability.
Summary by CodeRabbit