mirror of
https://github.com/ParisNeo/lollms-webui.git
synced 2025-04-10 04:09:55 +00:00
Enhanced query prepare
This commit is contained in:
Saifeddine ALOUI
parent
2e860ebb3b
commit
a3d8833694
269
api/__init__.py
269
api/__init__.py
@ -34,6 +34,7 @@ import re
|
||||
import string
|
||||
import requests
|
||||
from datetime import datetime
|
||||
from typing import List, Tuple
|
||||
|
||||
def terminate_thread(thread):
|
||||
if thread:
|
||||
@ -1194,76 +1195,222 @@ class LoLLMsAPPI(LollmsApplication):
|
||||
return cleaned_string
|
||||
|
||||
|
||||
def prepare_query(self, client_id, message_id=-1, is_continue=False):
|
||||
def prepare_query(self, client_id: str, message_id: int = -1, is_continue: bool = False, n_tokens: int = 0) -> Tuple[str, str, List[str]]:
|
||||
"""
|
||||
Prepares the query for the model.
|
||||
|
||||
Args:
|
||||
client_id (str): The client ID.
|
||||
message_id (int): The message ID. Default is -1.
|
||||
is_continue (bool): Whether the query is a continuation. Default is False.
|
||||
n_tokens (int): The number of tokens. Default is 0.
|
||||
|
||||
Returns:
|
||||
Tuple[str, str, List[str]]: The prepared query, original message content, and tokenized query.
|
||||
"""
|
||||
|
||||
# Get the list of messages
|
||||
messages = self.connections[client_id]["current_discussion"].get_messages()
|
||||
full_message_list = []
|
||||
|
||||
# Find the index of the message with the specified message_id
|
||||
message_index = -1
|
||||
for i, message in enumerate(messages):
|
||||
if message.id< message_id or (message_id==-1 and i<len(messages)-1):
|
||||
if message.content!='' and (message.message_type<=MSG_TYPE.MSG_TYPE_FULL_INVISIBLE_TO_USER.value and message.message_type!=MSG_TYPE.MSG_TYPE_FULL_INVISIBLE_TO_AI.value):
|
||||
full_message_list.append("\n"+self.config.discussion_prompt_separator+message.sender+": "+message.content.strip())
|
||||
else:
|
||||
if message.id == message_id:
|
||||
message_index = i
|
||||
break
|
||||
|
||||
link_text = "\n" #self.personality.link_text
|
||||
if not is_continue:
|
||||
full_message_list.append(self.config.discussion_prompt_separator +message.sender.strip().replace(":","")+": "+message.content.strip()+link_text+self.personality.ai_message_prefix.strip())
|
||||
else:
|
||||
full_message_list.append(self.config.discussion_prompt_separator +message.sender.strip().replace(":","")+": "+message.content.strip())
|
||||
|
||||
|
||||
composed_messages = link_text.join(full_message_list)
|
||||
t = self.model.tokenize(composed_messages)
|
||||
cond_tk = self.model.tokenize(self.personality.personality_conditioning)
|
||||
n_t = len(t)
|
||||
n_cond_tk = len(cond_tk)
|
||||
max_prompt_stx_size = 3*int(self.config.ctx_size/4)
|
||||
if n_cond_tk+n_t>max_prompt_stx_size:
|
||||
nb_tk = max_prompt_stx_size-n_cond_tk
|
||||
composed_messages = self.model.detokenize(t[-nb_tk:])
|
||||
ASCIIColors.warning(f"Cropping discussion to fit context [using {nb_tk} tokens/{self.config.ctx_size}]")
|
||||
discussion_messages = composed_messages
|
||||
|
||||
|
||||
# Define current message
|
||||
current_message = messages[message_index]
|
||||
|
||||
# Build the conditionning text block
|
||||
conditionning = self.personality.personality_conditioning
|
||||
if self.config["override_personality_model_parameters"]:
|
||||
conditionning = conditionning+ "\n!@>user description:\nName:"+self.config["user_name"]+"\n"+self.config["user_description"]+"\n"
|
||||
|
||||
str_docs = ""
|
||||
|
||||
if self.config.use_discussions_history:
|
||||
if self.discussions_store is not None:
|
||||
pr = PromptReshaper("{{conditionning}}\n!@>document chunks:\n{{doc}}\n{{content}}")
|
||||
docs, sorted_similarities = self.discussions_store.recover_text(message.content, top_k=self.config.data_vectorization_nb_chunks)
|
||||
for doc, infos in zip(docs, sorted_similarities):
|
||||
str_docs+=f"discussion chunk:\ndiscussion title: {infos[0]}\nchunk content:{doc}"
|
||||
|
||||
|
||||
if len(self.personality.files)>0 and self.personality.vectorizer:
|
||||
docs, sorted_similarities = self.personality.vectorizer.recover_text(message.content, top_k=self.config.data_vectorization_nb_chunks)
|
||||
# Check if there are document files to add to the prompt
|
||||
documentation = ""
|
||||
if len(self.personality.files) > 0 and self.personality.vectorizer:
|
||||
docs, sorted_similarities = self.personality.vectorizer.recover_text(current_message.content, top_k=self.config.data_vectorization_nb_chunks)
|
||||
for doc, infos in zip(docs, sorted_similarities):
|
||||
str_docs+=f"document chunk:\nchunk path: {infos[0]}\nchunk content:{doc}"
|
||||
documentation += f"document chunk:\nchunk path: {infos[0]}\nchunk content:{doc}"
|
||||
|
||||
if str_docs!="":
|
||||
pr = PromptReshaper("{{conditionning}}\n!@>document chunks:\n{{doc}}\n{{content}}")
|
||||
discussion_messages = pr.build({
|
||||
"doc":str_docs,
|
||||
"conditionning":conditionning,
|
||||
"content":discussion_messages
|
||||
}, self.model.tokenize, self.model.detokenize, self.config.ctx_size-self.config.min_n_predict, place_holders_to_sacrifice=["content"])
|
||||
# Check if there is discussion history to add to the prompt
|
||||
history = ""
|
||||
if self.config.use_discussions_history and self.discussions_store is not None:
|
||||
docs, sorted_similarities = self.discussions_store.recover_text(current_message.content, top_k=self.config.data_vectorization_nb_chunks)
|
||||
for doc, infos in zip(docs, sorted_similarities):
|
||||
history += f"discussion chunk:\ndiscussion title: {infos[0]}\nchunk content:{doc}"
|
||||
|
||||
# Tokenize the conditionning text and calculate its number of tokens
|
||||
tokens_conditionning = self.model.tokenize(conditionning)
|
||||
n_cond_tk = len(tokens_conditionning)
|
||||
|
||||
# Tokenize the documentation text and calculate its number of tokens
|
||||
if len(documentation)>0:
|
||||
tokens_documentation = self.model.tokenize(documentation)
|
||||
n_doc_tk = len(tokens_documentation)
|
||||
else:
|
||||
pr = PromptReshaper("{{conditionning}}\n{{content}}")
|
||||
discussion_messages = pr.build({
|
||||
"conditionning":conditionning,
|
||||
"content":discussion_messages
|
||||
}, self.model.tokenize, self.model.detokenize, self.config.ctx_size-self.config.min_n_predict, place_holders_to_sacrifice=["content"])
|
||||
# remove extra returns
|
||||
discussion_messages = self.clean_string(discussion_messages)
|
||||
tokens = self.model.tokenize(discussion_messages)
|
||||
if self.config["debug"]:
|
||||
ASCIIColors.yellow(discussion_messages)
|
||||
ASCIIColors.info(f"prompt size:{len(tokens)} tokens")
|
||||
tokens_documentation = []
|
||||
n_doc_tk = 0
|
||||
|
||||
# Tokenize the history text and calculate its number of tokens
|
||||
if len(history)>0:
|
||||
tokens_history = self.model.tokenize(history)
|
||||
n_history_tk = len(tokens_history)
|
||||
else:
|
||||
tokens_history = []
|
||||
n_history_tk = 0
|
||||
|
||||
# Calculate the total number of tokens between conditionning, documentation, and history
|
||||
total_tokens = n_cond_tk + n_doc_tk + n_history_tk
|
||||
|
||||
# Calculate the available space for the messages
|
||||
available_space = self.config.ctx_size - n_tokens - total_tokens
|
||||
|
||||
# Raise an error if the available space is 0 or less
|
||||
if available_space<1:
|
||||
raise Exception("Not enough space in context!!")
|
||||
|
||||
# Accumulate messages until the cumulative number of tokens exceeds available_space
|
||||
tokens_accumulated = 0
|
||||
|
||||
|
||||
# Initialize a list to store the full messages
|
||||
full_message_list = []
|
||||
# If this is not a continue request, we add the AI prompt
|
||||
if not is_continue:
|
||||
message_tokenized = self.model.tokenize(
|
||||
"\n" +self.personality.ai_message_prefix.strip()
|
||||
)
|
||||
full_message_list.append(message_tokenized)
|
||||
# Update the cumulative number of tokens
|
||||
tokens_accumulated += len(message_tokenized)
|
||||
|
||||
|
||||
# Accumulate messages starting from message_index
|
||||
for i in range(message_index, -1, -1):
|
||||
message = messages[i]
|
||||
|
||||
# Check if the message content is not empty and visible to the AI
|
||||
if message.content != '' and (
|
||||
message.message_type <= MSG_TYPE.MSG_TYPE_FULL_INVISIBLE_TO_USER.value and message.message_type != MSG_TYPE.MSG_TYPE_FULL_INVISIBLE_TO_AI.value):
|
||||
|
||||
# Tokenize the message content
|
||||
message_tokenized = self.model.tokenize(
|
||||
"\n" + self.config.discussion_prompt_separator + message.sender + ": " + message.content.strip())
|
||||
|
||||
# Check if adding the message will exceed the available space
|
||||
if tokens_accumulated + len(message_tokenized) > available_space:
|
||||
break
|
||||
|
||||
# Add the tokenized message to the full_message_list
|
||||
full_message_list.insert(0, message_tokenized)
|
||||
|
||||
# Update the cumulative number of tokens
|
||||
tokens_accumulated += len(message_tokenized)
|
||||
|
||||
# Build the final discussion messages by detokenizing the full_message_list
|
||||
discussion_messages = ""
|
||||
for message_tokens in full_message_list:
|
||||
discussion_messages += self.model.detokenize(message_tokens)
|
||||
|
||||
# Build the final prompt by concatenating the conditionning and discussion messages
|
||||
prompt_data = conditionning + discussion_messages
|
||||
|
||||
# Tokenize the prompt data
|
||||
tokens = self.model.tokenize(prompt_data)
|
||||
|
||||
# if this is a debug then show prompt construction details
|
||||
if self.config["debug"]:
|
||||
ASCIIColors.bold("CONDITIONNING")
|
||||
ASCIIColors.yellow(conditionning)
|
||||
ASCIIColors.bold("DOC")
|
||||
ASCIIColors.yellow(documentation)
|
||||
ASCIIColors.bold("HISTORY")
|
||||
ASCIIColors.yellow(history)
|
||||
ASCIIColors.bold("DISCUSSION")
|
||||
ASCIIColors.hilight(discussion_messages,"!@>",ASCIIColors.color_yellow,ASCIIColors.color_bright_red,False)
|
||||
ASCIIColors.bold("Final prompt")
|
||||
ASCIIColors.hilight(prompt_data,"!@>",ASCIIColors.color_yellow,ASCIIColors.color_bright_red,False)
|
||||
ASCIIColors.info(f"prompt size:{len(tokens)} tokens")
|
||||
ASCIIColors.info(f"available space after doc and history:{available_space} tokens")
|
||||
|
||||
# Return the prepared query, original message content, and tokenized query
|
||||
return prompt_data, current_message.content, tokens
|
||||
|
||||
|
||||
"""
|
||||
def prepare_query(self, client_id, message_id=-1, is_continue=False):
|
||||
messages = self.connections[client_id]["current_discussion"].get_messages()
|
||||
full_message_list = []
|
||||
for i, message in enumerate(messages):
|
||||
if message.id< message_id or (message_id==-1 and i<len(messages)-1):
|
||||
if message.content!='' and (message.message_type<=MSG_TYPE.MSG_TYPE_FULL_INVISIBLE_TO_USER.value and message.message_type!=MSG_TYPE.MSG_TYPE_FULL_INVISIBLE_TO_AI.value):
|
||||
full_message_list.append("\n"+self.config.discussion_prompt_separator+message.sender+": "+message.content.strip())
|
||||
else:
|
||||
break
|
||||
|
||||
link_text = "\n" #self.personality.link_text
|
||||
if not is_continue:
|
||||
full_message_list.append(self.config.discussion_prompt_separator +message.sender.strip().replace(":","")+": "+message.content.strip()+link_text+self.personality.ai_message_prefix.strip())
|
||||
else:
|
||||
full_message_list.append(self.config.discussion_prompt_separator +message.sender.strip().replace(":","")+": "+message.content.strip())
|
||||
|
||||
|
||||
composed_messages = link_text.join(full_message_list)
|
||||
t = self.model.tokenize(composed_messages)
|
||||
cond_tk = self.model.tokenize(self.personality.personality_conditioning)
|
||||
n_t = len(t)
|
||||
n_cond_tk = len(cond_tk)
|
||||
max_prompt_stx_size = 3*int(self.config.ctx_size/4)
|
||||
if n_cond_tk+n_t>max_prompt_stx_size:
|
||||
nb_tk = max_prompt_stx_size-n_cond_tk
|
||||
composed_messages = self.model.detokenize(t[-nb_tk:])
|
||||
ASCIIColors.warning(f"Cropping discussion to fit context [using {nb_tk} tokens/{self.config.ctx_size}]")
|
||||
discussion_messages = composed_messages
|
||||
|
||||
|
||||
conditionning = self.personality.personality_conditioning
|
||||
if self.config["override_personality_model_parameters"]:
|
||||
conditionning = conditionning+ "\n!@>user description:\nName:"+self.config["user_name"]+"\n"+self.config["user_description"]+"\n"
|
||||
|
||||
str_docs = ""
|
||||
|
||||
if self.config.use_discussions_history:
|
||||
if self.discussions_store is not None:
|
||||
pr = PromptReshaper("{{conditionning}}\n!@>document chunks:\n{{doc}}\n{{content}}")
|
||||
docs, sorted_similarities = self.discussions_store.recover_text(message.content, top_k=self.config.data_vectorization_nb_chunks)
|
||||
for doc, infos in zip(docs, sorted_similarities):
|
||||
str_docs+=f"discussion chunk:\ndiscussion title: {infos[0]}\nchunk content:{doc}"
|
||||
|
||||
|
||||
if len(self.personality.files)>0 and self.personality.vectorizer:
|
||||
docs, sorted_similarities = self.personality.vectorizer.recover_text(message.content, top_k=self.config.data_vectorization_nb_chunks)
|
||||
for doc, infos in zip(docs, sorted_similarities):
|
||||
str_docs+=f"document chunk:\nchunk path: {infos[0]}\nchunk content:{doc}"
|
||||
|
||||
if str_docs!="":
|
||||
pr = PromptReshaper("{{conditionning}}\n!@>document chunks:\n{{doc}}\n{{content}}")
|
||||
discussion_messages = pr.build({
|
||||
"doc":str_docs,
|
||||
"conditionning":conditionning,
|
||||
"content":discussion_messages
|
||||
}, self.model.tokenize, self.model.detokenize, self.config.ctx_size-self.config.min_n_predict, place_holders_to_sacrifice=["content"])
|
||||
else:
|
||||
pr = PromptReshaper("{{conditionning}}\n{{content}}")
|
||||
discussion_messages = pr.build({
|
||||
"conditionning":conditionning,
|
||||
"content":discussion_messages
|
||||
}, self.model.tokenize, self.model.detokenize, self.config.ctx_size-self.config.min_n_predict, place_holders_to_sacrifice=["content"])
|
||||
# remove extra returns
|
||||
discussion_messages = self.clean_string(discussion_messages)
|
||||
tokens = self.model.tokenize(discussion_messages)
|
||||
if self.config["debug"]:
|
||||
ASCIIColors.yellow(discussion_messages)
|
||||
ASCIIColors.info(f"prompt size:{len(tokens)} tokens")
|
||||
|
||||
return discussion_messages, message.content, tokens
|
||||
|
||||
"""
|
||||
|
||||
return discussion_messages, message.content, tokens
|
||||
|
||||
def get_discussion_to(self, client_id, message_id=-1):
|
||||
messages = self.connections[client_id]["current_discussion"].get_messages()
|
||||
@ -1529,7 +1676,7 @@ class LoLLMsAPPI(LollmsApplication):
|
||||
self.nb_received_tokens = 0
|
||||
self.start_time = datetime.now()
|
||||
if self.model is not None:
|
||||
ASCIIColors.info(f"warmup for generating {n_predict} tokens")
|
||||
ASCIIColors.info(f"warmup for generating up to {n_predict} tokens")
|
||||
if self.config["override_personality_model_parameters"]:
|
||||
output = self.model.generate(
|
||||
prompt,
|
||||
|
||||
@ -12,4 +12,6 @@ So, whether you're an aspiring artist or a seasoned creator looking to explore t
|
||||
Before we dive into the exciting world of artbot, let's take a moment to understand the terms of use. Artbot is an incredible AI-powered tool that utilizes an LLM (Large Language Model) to generate positive and negative prompts, which are then fed to an image generation model. Additionally, artbot has the ability to view your images and convert them into text descriptions using Blip, allowing the LLM to comprehend them better. What's fascinating is that artbot can also engage in conversations with you, enabling you to build upon vague ideas and transform them into something truly beautiful and original.
|
||||
Now, it's essential to be transparent about the collaborative nature of working with AI. When presenting your artwork created with artbot, it's crucial to explicitly acknowledge the collaboration with AI and not claim it as your own manual production. The purpose of these tools is to inspire and assist you in thinking creatively, ultimately enhancing your artistic abilities. Transparency is key, as flooding the world with AI-generated content can lead to unfair competition with individuals who dedicate time and care to manual work. Manual work should be valued and appreciated in its own right.
|
||||
Therefore, let's be honest and focus on valuing the ideas behind the artwork. When someone purchases your NFT, they should be aware that it is AI-generated, and the added value comes from your initiative and collaboration with artbot. By doing so, we ensure that the true worth of manual work is recognized, and the creative potential of AI is utilized responsibly.
|
||||
Now that we've set the stage, let's unleash our imagination and embark on an artistic journey with artbot. Get ready to witness the incredible possibilities and showcase your unique artistic vision to the world through NFTs on opensea!
|
||||
Now that we've set the stage, let's unleash our imagination and embark on an artistic journey with artbot. Get ready to witness the incredible possibilities and showcase your unique artistic vision to the world through NFTs on opensea!
|
||||
|
||||
|
||||
|
||||
@ -1 +1 @@
|
||||
Subproject commit c7a70a7473a3ba6052cbc44b041524bcdbdf8eb6
|
||||
Subproject commit 893f20f4db2e82b5891e9bd30a0d2facd9d98280
|
||||
Loading…
x
Reference in New Issue
Block a user