projectmoon
/
open-webui-filters
1
0
Fork 0
Code Issues 3 Pull Requests Packages Projects Releases Wiki Activity

Checkpoint summarization filter.

This commit is contained in:
projectmoon 2024-07-25 00:55:58 +02:00
parent 5649d4582d
commit 8f00b8ebd8
2 changed files with 873 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

78
README.md
View File

@ -6,6 +6,8 @@ My collection of OpenWebUI Filters.
So far:
- **Checkpoint Summarization Filter:** A work-in-progress replacement
for the narrative memory filter for more generalized use cases.
- **Memory Filter:** A basic narrative memory filter intended for
long-form storytelling/roleplaying scenarios. Intended as a proof
of concept/springboard for more advanced narrative memory.
@ -14,8 +16,83 @@ So far:
- **Output Sanitization Filter:** Remove words, phrases, or
characters from the start of model replies.
## Checkpoint Sumarization Filter
A new filter for managing context use by summarizing previous parts of
the chat as the conversation continues. Designed for both general
chats and narrative/roleplay use. Work in progress.
### Configuration
There are currently 4 settings:
- **Summarizer Model:** The model used to summarize the conversation
as the chat continues. This must be a base model.
- **Large Context Summarizer Model:** If large context summarization
is turned on, use this model for summarizing huge contexts.
- **Summarize Large Contexts:** If enabled, the filter will attempt
to load the entire context into the large summarizer model for
creating an initial checkpoint of an existing conversation.
- **Wiggle Room:** This is the amount of 'wiggle room' for estimating
a context shift. This number is subtracted from `num_ctx` for the
purposes of determining whether or not a context shift has
occurred.
### Usage
In general, you should only need to specify the summarizer model and
enable the filter on the OpenWebUI models that you want it to work on.
Or even enable it globally. The filter works best when used from a new
conversation, but it does have the (currently limited) ability to deal
with existing conversations.
- When the filter detects a context shift in the conversation, it
will summarize the pre-existing context. After that, the summary
is appended to the system prompt, and old messages before
summarization are dropped.
- When the filter detects the next context shift, this process is
repeated, and a new summarization checkpoint is created. And so
on.
If the filter is used in an existing conversation, it will summarize
on the first time that it detects a context shift in the conversation:
- If there are enough messages that the conversation is considered
"big," and large context summarization is **disabled**, all but the
last 4 messages will be dropped to form the summary.
- If the conversation is considered "big," and large context
summarization is **enabled**, then the large context model will be
loaded to do the summarization, and the **entire conversation**
will be given to it.
#### User Commands
There are some basic commands the user can use to interact with the
filter in a conversation:
- `!nuke`: Deletes all summary checkpoints in the chat, and the
filter will attempt to summarize from scratch the next time it
detects a context shift.
### Limitations
There are some limitations to be aware of:
- If you enable large context summarization, you need to make sure
your system is capable of loading and summarizing an entire
conversation.
- Handling of branching conversations and regenerated responses is
currently rather messy. It will kind of work. There are some plans
to improve this.
- If large context summarization is disabled, pre-existing large
conversations will only summarize the previous 4 messages when the
first summarization is detected.
- The filter only loads the most recent summary, and thus the AI
might "forget" much older information.
## Memory Filter
__Superseded By: [Checkpoint Summarization Filter][checkpoint-filter]__
Super hacky, very basic automatic narrative memory filter for
OpenWebUI, that may or may not actually enhance narrative generation!
@ -154,3 +231,4 @@ aware how this might affect your OpenWebUI deployment, if you are
deploying OpenWebUI in a public environment!
[agpl]: https://www.gnu.org/licenses/agpl-3.0.en.html
[checkpoint-filter]: #checkpoint-summarization-filter

795
checkpoint_summary_filter.py Normal file
View File

@ -0,0 +1,795 @@
"""
title: Checkpoint Summary Filter
author: projectmoon
author_url: https://git.agnos.is/projectmoon/open-webui-filters
version: 0.1.0
license: AGPL-3.0+
required_open_webui_version: 0.3.9
"""
# Documentation: https://git.agnos.is/projectmoon/open-webui-filters
# System imports
import asyncio
import hashlib
import uuid
import json
import re
import logging
from typing import Optional, List, Dict, Callable, Any, NewType, Tuple, Awaitable, ClassVar
from typing_extensions import TypedDict, NotRequired
from collections import deque
# Libraries available to OpenWebUI
from pydantic import BaseModel as PydanticBaseModel, Field
import chromadb
from chromadb import Collection as ChromaCollection
from chromadb.api.types import Document as ChromaDocument
# OpenWebUI imports
from config import CHROMA_CLIENT
from apps.rag.main import app as rag_app
from apps.ollama.main import app as ollama_app
from apps.ollama.main import show_model_info, ModelNameForm
from utils.misc import get_last_user_message, get_last_assistant_message
from main import generate_chat_completions
from apps.webui.models.chats import Chats
from apps.webui.models.models import Models
from apps.webui.models.users import Users
# Embedding (not yet used)
EMBEDDING_FUNCTION = rag_app.state.EMBEDDING_FUNCTION
EmbeddingFunc = NewType('EmbeddingFunc', Callable[[str], List[Any]])
# Prompts
SUMMARIZER_PROMPT = """
### Main Instructions
You are a chat conversation summarizer. Your task is to summarize the given
portion of an ongoing conversation. First, determine if the conversation is
a regular chat between the user and the assistant, or if the conversation is
part of a story or role-playing session.
Summarize the important parts of the given chat between the user and the
assistant. Limit your summary to one paragraph. Make sure your summary is
detailed. Write the summary as if you are summarizing part of a larger
conversation.
### Regular Chat
If the conversation is a regular chat, write your summary referring to the
ongoing conversation as a chat. Refer to the user and the assistant as user
and assistant. Do not refer to yourself as the assistant.
### Story or Role-Playing Session
If the conversation is a story or role-playing session, write your summary
referring to the conversation as an ongoing story. Do not refer to the user
or assistant in your summary. Only use the names of the characters, places,
and events in the story.
""".replace("\n", " ").strip()
class Message(TypedDict):
id: NotRequired[str]
role: str
content: str
class MessageInsertMetadata(TypedDict):
role: str
chapter: str
class MessageInsert(TypedDict):
message_id: str
content: str
metadata: MessageInsertMetadata
embeddings: List[Any]
class BaseModel(PydanticBaseModel):
class Config:
arbitrary_types_allowed = True
class SummarizerResponse(BaseModel):
summary: str
class Summarizer(BaseModel):
messages: List[dict]
model: str
prompt: str = SUMMARIZER_PROMPT
async def summarize(self) -> Optional[SummarizerResponse]:
sys_message: Message = { "role": "system", "content": SUMMARIZER_PROMPT }
user_message: Message = {
"role": "user",
"content": "Make a detailed summary of everything up to this point."
}
messages = [sys_message] + self.messages + [user_message]
request = {
"model": self.model,
"messages": messages,
"stream": False,
"keep_alive": "10s"
}
resp = await generate_chat_completions(request)
if "choices" in resp and len(resp["choices"]) > 0:
content: str = resp["choices"][0]["message"]["content"]
return SummarizerResponse(summary=content)
else:
return None
class Checkpoint(BaseModel):
# chat id
chat_id: str
# the message ID this checkpoint was created from.
message_id: str
# index of the message in the message input array. in the inlet
# function, we do not have access to incoming message ids for some
# reason. used as a fallback to drop old context when
message_index: int = 0
# the "slug", or chain of messages, that led to this point.
slug: str
# actual summary of messages.
summary: str
# if we try to put a type hint on this, it gets mad.
@staticmethod
def from_json(obj: dict):
try:
return Checkpoint(
chat_id=obj["chat_id"],
message_id=obj["message_id"],
message_index=obj["message_index"],
slug=obj["slug"],
summary=obj["summary"]
)
except:
return None
def to_json(self) -> str:
return self.model_dump_json()
class Checkpointer(BaseModel):
"""Manages summary checkpoints in a single chat."""
chat_id: str
summarizer_model: str = ""
chroma_client: chromadb.ClientAPI
messages: List[dict]=[] # stripped set of messages
full_messages: List[dict]=[] # all the messages
embedding_func: EmbeddingFunc=(lambda a: 0)
collection_name: ClassVar[str] = "chat_checkpoints"
def _get_collection(self) -> ChromaCollection:
return self.chroma_client.get_or_create_collection(
name=Checkpointer.collection_name
)
def _insert_checkpoint(self, checkpoint: Checkpoint):
coll = self._get_collection()
checkpoint_doc = checkpoint.to_json()
# Insert the checkpoint itself with slug as ID.
coll.upsert(
ids=[checkpoint.slug],
documents=[checkpoint_doc],
metadatas=[{ "chat_id": self.chat_id, "type": "checkpoint" }],
embeddings=[self.embedding_func(checkpoint_doc)]
)
# Update the chat info doc for this chat.
coll.upsert(
ids=[self.chat_id],
documents=[json.dumps({ "current_checkpoint": checkpoint.slug })],
embeddings=[self.embedding_func(self.chat_id)]
)
def _calculate_slug(self) -> Optional[str]:
if len(self.messages) == 0:
return None
message_ids = [msg["id"] for msg in reversed(self.messages)]
slug = "|".join(message_ids)
return hashlib.sha256(slug.encode()).hexdigest()
def _get_state(self):
resp = self._get_collection().get(ids=[self.chat_id], include=["documents"])
state: dict = (json.loads(resp["documents"][0])
if resp["documents"] and len(resp["documents"]) > 0
else { "current_checkpoint": None })
return state
def _find_message_index(self, message_id: str) -> Optional[int]:
for idx, message in enumerate(self.full_messages):
if message["id"] == message_id:
return idx
return None
def nuke_checkpoints(self):
"""Delete all checkpoints for this chat."""
coll = self._get_collection()
checkpoints = coll.get(
include=["documents"],
where={"chat_id": self.chat_id}
)
self._get_collection().delete(
ids=[self.chat_id] + checkpoints["ids"]
)
async def create_checkpoint(self) -> str:
summarizer = Summarizer(model=self.summarizer_model, messages=self.messages)
resp = await summarizer.summarize()
if resp:
slug = self._calculate_slug()
checkpoint_message = self.messages[-1]
checkpoint_index = self._find_message_index(checkpoint_message["id"])
checkpoint = Checkpoint(
chat_id = self.chat_id,
slug = self._calculate_slug(),
message_id = checkpoint_message["id"],
message_index = checkpoint_index,
summary = resp.summary
)
self._insert_checkpoint(checkpoint)
return slug
def get_checkpoint(self, slug: Optional[str]) -> Optional[Checkpoint]:
if not slug:
return None
resp = self._get_collection().get(ids=[slug], include=["documents"])
checkpoint = (resp["documents"][0]
if resp["documents"] and len(resp["documents"]) > 0
else None)
if checkpoint:
return Checkpoint.from_json(json.loads(checkpoint))
else:
return None
def get_current_checkpoint(self) -> Optional[Checkpoint]:
state = self._get_state()
return self.get_checkpoint(state["current_checkpoint"])
#########################
# Utilities
#########################
class SessionInfo(BaseModel):
chat_id: str
message_id: str
session_id: str
def extract_session_info(event_emitter) -> Optional[SessionInfo]:
"""The latest innovation in hacky workarounds."""
try:
info = event_emitter.__closure__[0].cell_contents
return SessionInfo(
chat_id=info["chat_id"],
message_id=info["message_id"],
session_id=info["session_id"]
)
except:
return None
def predicted_token_use(messages) -> int:
"""Parse most recent message to calculate estimated token use."""
if len(self.messages == 0):
return 0
# Naive assumptions:
# - 1 word = 1 token.
# - 1 period, comma, or colon = 1 token
message = messages[-1]
return len(list(filter(None, re.split(r"\s|(;)|(,)|(\.)|(:)|\n", message))))
def is_big_convo(messages, num_ctx: int=8192) -> bool:
"""
Attempt to detect large pre-existing conversation by looking at
recent eval counts from messages and comparing against given
num_ctx. We check all messages for an eval count that goes above
the context limit. It doesn't matter where in the message list; if
it's somewhere in the middle, it means that there was a context
shift.
"""
for message in messages:
if "info" in message:
tokens_used = (message["info"]["eval_count"] +
message["info"]["prompt_eval_count"])
else:
tokens_used = 0
if tokens_used >= num_ctx:
return True
return False
def hit_context_limit(
messages,
num_ctx: int=8192,
wiggle_room: int=1000
) -> Tuple[bool, int]:
"""
Determine if we've hit the context limit, within some reasonable
estimation. We have a defined 'wiggle room' that is subtracted
from the num_ctx parameter, in order to capture near-filled
contexts. We do it this way because we're summarizing on output,
rather than before input (inlet function doesn't have enough
info).
"""
if len(messages) == 0:
return False, 0
last_message = messages[-1]
tokens_used = 0
if "info" in last_message:
tokens_used = (last_message["info"]["eval_count"] +
last_message["info"]["prompt_eval_count"])
if tokens_used >= (num_ctx - wiggle_room):
amount_over = tokens_used - num_ctx
amount_over = 0 if amount_over < 0 else amount_over
return True, amount_over
else:
return False, 0
def extract_base_model_id(model: dict) -> Optional[str]:
if "base_model_id" not in model["info"]:
return None
base_model_id = model["info"]["base_model_id"]
if not base_model_id:
base_model_id = model["id"]
return base_model_id
def extract_owu_model_param(model_obj: dict, param_name: str):
"""
Extract a parameter value from the DB definition of a model
that is based on another model.
"""
if not "params" in model_obj["info"]:
return None
params = model_obj["info"]["params"]
return params.get(param_name, None)
def extract_owu_base_model_param(base_model_id: str, param_name: str):
"""Extract a parameter value from the DB definition of an ollama base model."""
base_model = Models.get_model_by_id(base_model_id)
if not base_model:
return None
base_model.params = base_model.params.model_dump()
return base_model.params.get(param_name, None)
def extract_ollama_response_param(model: dict, param_name: str):
"""Extract a parameter value from ollama show API response."""
if "parameters" not in model:
return None
for line in model["parameters"].splitlines():
if line.startswith(param_name):
return line.lstrip(param_name).strip()
return None
async def get_model_from_ollama(model_id: str, user_id) -> Optional[dict]:
"""Call ollama show API and return model information."""
curr_user = Users.get_user_by_id(user_id)
try:
return await show_model_info(ModelNameForm(name=model_id), user=curr_user)
except Exception as e:
print(f"Could not get model info: {e}")
return None
async def calculate_num_ctx(chat_id: str, user_id, model: dict) -> int:
"""
Attempt to discover the current num_ctx parameter in many
different ways.
"""
# first check the open-webui chat parameters.
chat = Chats.get_chat_by_id_and_user_id(chat_id, user_id)
if chat:
# this might look odd, but the chat field is a json blob of
# useful info.
chat = json.loads(chat.chat)
if "params" in chat and "num_ctx" in chat["params"]:
return chat["params"]["num_ctx"]
# then check open web ui model def
num_ctx = extract_owu_model_param(model, "num_ctx")
if num_ctx:
return num_ctx
# then check open web ui base model def.
base_model_id = extract_base_model_id(model)
if not base_model_id:
# fall back to default in case of weirdness.
return 2048
num_ctx = extract_owu_base_model_param(base_model_id, "num_ctx")
if num_ctx:
return num_ctx
# THEN check ollama directly.
base_model = await get_model_from_ollama(base_model_id, user_id)
num_ctx = extract_ollama_response_param(base_model, "num_ctx")
if num_ctx:
return num_ctx
# finally, return default.
return 2048
class Filter:
class Valves(BaseModel):
def summarizer_model(self, body):
if self.summarizer_model_id == "":
return extract_base_model_id(body["model"])
else:
return self.summarizer_model_id
summarize_large_contexts: bool = Field(
default=False,
description=(
f"Whether or not to use a large context model to summarize large "
f"pre-existing conversations."
)
)
wiggle_room: int = Field(
default=1000,
description=(
"Amount of token 'wiggle room' for estimating when a context shift occurs. "
"Subtracted from num_ctx when checking if summarization is needed."
)
)
summarizer_model_id: str = Field(
default="",
description="Model used to summarize the conversation. Must be a base model.",
)
large_summarizer_model_id: str = Field(
default="",
description=(
"Model used to summarize large pre-existing contexts. "
"Must be a base model with a context size large enough "
"to fit the conversation."
)
)
pass
class UserValves(BaseModel):
pass
def __init__(self):
self.valves = self.Valves()
pass
def load_current_chat(self) -> dict:
# the chat property of the model is the json blob that holds
# all the interesting stuff
chat = (Chats
.get_chat_by_id_and_user_id(self.session_info.chat_id, self.user["id"])
.chat)
return json.loads(chat)
def get_messages_for_checkpointing(self, messages, num_ctx, last_checkpointed_id):
"""
Assemble list of messages to checkpoint, based on current
state and valve settings.
"""
message_chain = deque()
for message in reversed(messages):
if message["id"] == last_checkpointed_id:
break
message_chain.appendleft(message)
message_chain = list(message_chain) # the lazy way
# now we check if we are a big conversation, and if valve
# settings allow that kind of summarization.
summarizer_model = self.valves.summarizer_model
if is_big_convo(messages, num_ctx) and not self.valves.summarize_large_contexts:
# must summarize using small model. for now, drop to last
# N messages.
print((
"Dropping all but last 4 messages to summarize "
"large convo without large model."
))
message_chain = message_chain[-4:]
return message_chain
async def create_checkpoint(
self,
messages: List[dict],
last_checkpointed_id: Optional[str]=None,
num_ctx: int=8192
):
if len(messages) == 0:
return
print(f"[{self.session_info.chat_id}] Detected context shift. Summarizing.")
await self.set_summarizing_status(done=False)
last_message = messages[-1] # should check for role = assistant
curr_message_id: Optional[str] = (
last_message["id"] if last_message else None
)
if not curr_message_id:
return
# strip messages down to what is in the current checkpoint.
message_chain = self.get_messages_for_checkpointing(
messages, num_ctx, last_checkpointed_id
)
# we should now have a list of messages that is just within
# the current context limit.
summarizer_model = self.valves.summarizer_model_id
if is_big_convo(message_chain, num_ctx) and self.valves.summarize_large_contexts:
print("Summarizing LARGE context!")
summarizer_model = self.valves.large_summarizer_model_id
checkpointer = Checkpointer(
chat_id=self.session_info.chat_id,
summarizer_model=summarizer_model,
chroma_client=CHROMA_CLIENT,
full_messages=messages,
messages=message_chain
)
try:
slug = await checkpointer.create_checkpoint()
await self.set_summarizing_status(done=True)
print(("Summarization checkpoint created in chat "
f"'{self.session_info.chat_id}': {slug}"))
except Exception as e:
print(f"Error creating summary: {str(e)}")
await self.set_summarizing_status(
done=True, message=f"Error summarizing: {str(e)}"
)
def update_chat_with_checkpoint(self, messages: List[dict], checkpoint: Checkpoint):
if len(messages) < checkpoint.message_index:
# do not mess with anything if the index doesn't even
# exist anymore. need a new checkpoint.
return messages
# proceed with altering the system prompt. keep system prompt,
# if it's there, and add summary to it. summary will become
# system prompt if there is no system prompt.
convo_messages = [
message for message in messages if message.get("role") != "system"
]
system_prompt = next(
(message for message in messages if message.get("role") == "system"), None
)
summary_message = f"Summary of conversation so far:\n\n{checkpoint.summary}"
if system_prompt:
system_prompt["content"] += f"\n\n{summary_message}"
else:
system_prompt = { "role": "system", "content": summary_message }
# drop old messages, reapply system prompt.
messages = self.apply_checkpoint(checkpoint, messages)
return [system_prompt] + messages
async def send_message(self, message: str):
await self.event_emitter({
"type": "status",
"data": {
"description": message,
"done": True,
},
})
async def set_summarizing_status(self, done: bool, message: Optional[str]=None):
if not self.event_emitter:
return
if not done:
description = (
"Summarizing conversation due to reaching context limit (do not reply yet)."
)
else:
description = (
"Summarization complete (you may now reply)."
)
if message:
description = message
await self.event_emitter({
"type": "status",
"data": {
"description": description,
"done": done,
},
})
def apply_checkpoint(
self, checkpoint: Checkpoint, messages: List[dict]
) -> List[dict]:
"""
Possibly shorten the message context based on a checkpoint.
This works two ways: if the messages have IDs (outlet
filter), split by message ID (very reliable). Otherwise,
attempt to split by on the recorded message index (inlet
filter; not very reliable).
"""
# first attempt to drop everything before the checkpointed
# message id.
split_point = 0
for idx, message in enumerate(messages):
if "id" in message and message["id"] == checkpoint.message_id:
split_point = idx
break
# if we can't find the ID to split on, fall back to message
# index if possible. this can happen during message
# regeneration, for example. or if we're called from the inlet
# filter, which doesn't have access to message ids.
if split_point == 0 and checkpoint.message_index <= len(messages):
split_point = checkpoint.message_index
orig = len(messages)
messages = messages[split_point:]
print((f"[{self.session_info.chat_id}] Dropped context to {len(messages)} "
f"messages (from {orig})"))
return messages
async def handle_nuke(self, body):
checkpointer = Checkpointer(
chat_id=self.session_info.chat_id,
chroma_client=CHROMA_CLIENT
)
checkpointer.nuke_checkpoints()
await self.send_message("Deleted all checkpoint for chat.")
body["messages"][-1]["content"] = (
"Respond ony with: 'Deleted all checkpoint for chat.'"
)
body["messages"] = body["messages"][-1:]
return body
async def outlet(
self,
body: dict,
__user__: Optional[dict],
__model__: Optional[dict],
__event_emitter__: Callable[[Any], Awaitable[None]],
) -> dict:
# Useful things to have around.
self.user = __user__
self.model = __model__
self.session_info = extract_session_info(__event_emitter__)
self.event_emitter = __event_emitter__
self.summarizer_model_id = self.valves.summarizer_model(body)
# global filters apply to requests coming in through proxied
# API. If we're not an OpenWebUI chat, abort mission.
if not self.session_info:
return body
if not self.model or self.modle["owned_by"] != "ollama":
return body
messages = body["messages"]
num_ctx = await calculate_num_ctx(
chat_id=self.session_info.chat_id,
user_id=self.user["id"],
model=self.model
)
# apply current checkpoint ONLY for purposes of calculating if
# we have hit num_ctx within current checkpoint.
checkpointer = Checkpointer(
chat_id=self.session_info.chat_id,
chroma_client=CHROMA_CLIENT
)
checkpoint = checkpointer.get_current_checkpoint()
messages_for_ctx_check = (self.apply_checkpoint(checkpoint, messages)
if checkpoint else messages)
hit_limit, amount_over = hit_context_limit(
messages=messages_for_ctx_check,
num_ctx=num_ctx,
wiggle_room=self.valves.wiggle_room
)
if hit_limit:
# we need the FULL message list to do proper summarizing,
# because we might be summarizing a hug context.
await self.create_checkpoint(
messages=messages,
num_ctx=num_ctx,
last_checkpointed_id=checkpoint.message_id if checkpoint else None
)
print(f"[{self.session_info.chat_id}] Done checking for summarization")
return body
async def inlet(
self,
body: dict,
__user__: Optional[dict],
__model__: Optional[dict],
__event_emitter__: Callable[[Any], Awaitable[None]]
) -> dict:
# Useful properties to have around.
self.user = __user__
self.model = __model__
self.session_info = extract_session_info(__event_emitter__)
self.event_emitter = __event_emitter__
self.summarizer_model_id = self.valves.summarizer_model(body)
# global filters apply to requests coming in through proxied
# API. If we're not an OpenWebUI chat, abort mission.
if not self.session_info:
return body
if not self.model or self.modle["owned_by"] != "ollama":
return body
# super basic external command handling (delete checkpoints).
user_msg = get_last_user_message(body["messages"])
if user_msg and user_msg == "!nuke":
return await self.handle_nuke(body)
# apply current checkpoint to the chat: adds most recent
# summary to system prompt, and drops all messages before the
# checkpoint.
checkpointer = Checkpointer(
chat_id=self.session_info.chat_id,
chroma_client=CHROMA_CLIENT
)
checkpoint = checkpointer.get_current_checkpoint()
if checkpoint:
print((
f"Using checkpoint {checkpoint.slug} for "
f"conversation {self.session_info.chat_id}"
))
body["messages"] = self.update_chat_with_checkpoint(body["messages"], checkpoint)
return body