Skip to content

Commit

Permalink
Merge pull request #29 from log10-io/inception-prompts
Browse files Browse the repository at this point in the history 
Customizable inception prompts for agents
  • Loading branch information
@nqn
nqn authored Jun 30, 2023
2 parents 05512d2 + 9b0a58b commit 5312e05
Show file tree
Hide file tree
Showing 3 changed files with 109 additions and 48 deletions.
153 changes: 107 additions & 46 deletions log10/agents/camel.py
View file
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
import logging
from dotenv import load_dotenv

load_dotenv()
from anthropic import HUMAN_PROMPT
from log10.utils import convert_history_to_claude
Expand All @@ -10,16 +11,49 @@
# Repeat word termination conditions
# https://github.com/lightaime/camel/blob/master/examples/ai_society/role_playing_multiprocess.py#L63
repeat_word_threshold = 4
repeat_word_list = [
"goodbye", "good bye", "thank", "bye", "welcome", "language model"
]
repeat_word_list = ["goodbye", "good bye", "thank", "bye", "welcome", "language model"]


def camel_agent(userRole, assistantRole, taskPrompt, model, summary_model, maxTurns, module, hparams):
generator = camel_agent_generator(userRole, assistantRole, taskPrompt, model, summary_model, maxTurns, module, hparams)
def camel_agent(
userRole,
assistantRole,
taskPrompt,
model,
summary_model,
maxTurns,
module,
hparams,
userPrompt=None,
assistantPrompt=None,
):
generator = camel_agent_generator(
userRole,
userPrompt,
assistantRole,
assistantPrompt,
taskPrompt,
model,
summary_model,
maxTurns,
module,
hparams,
)
*_, last = generator
return last

def camel_agent_generator(userRole, assistantRole, taskPrompt, model, summary_model, maxTurns, module, hparams):

def camel_agent_generator(
userRole,
userPrompt,
assistantRole,
assistantPrompt,
taskPrompt,
model,
summary_model,
maxTurns,
module,
hparams,
):
try:
assistant_inception_prompt = f"""Never forget you are a {assistantRole} and I am a {userRole}. Never flip roles! Never instruct me!
We share a common interest in collaborating to successfully complete a task.
Expand All @@ -41,6 +75,9 @@ def camel_agent_generator(userRole, assistantRole, taskPrompt, model, summary_mo
<YOUR_SOLUTION> should be specific and provide preferable implementations and examples for task-solving.
Always end <YOUR_SOLUTION> with: Next request."""

if assistantPrompt is not None:
assistant_inception_prompt = assistantPrompt

user_inception_prompt = f"""Never forget you are a {userRole} and I am a {assistantRole}. Never flip roles! You will always instruct me.
We share a common interest in collaborating to successfully complete a task.
I must help you to complete the task.
Expand All @@ -67,89 +104,109 @@ def camel_agent_generator(userRole, assistantRole, taskPrompt, model, summary_mo
When the task is completed, you must only reply with a single word <CAMEL_TASK_DONE>.
Never say <CAMEL_TASK_DONE> unless my responses have solved your task."""

assistant_prompt = f'ASSISTANT PROMPT: \n {assistant_inception_prompt}'
user_prompt = f'USER PROMPT: \n {user_inception_prompt}'
if userPrompt is not None:
user_inception_prompt = userPrompt

assistant_prompt = f"ASSISTANT PROMPT: \n {assistant_inception_prompt}"
user_prompt = f"USER PROMPT: \n {user_inception_prompt}"

assistant_messages = [{"role": "system", "content": assistant_prompt},
{"role": "user", "content": assistant_prompt}]
user_messages = [{"role": "system", "content": user_prompt},
{"role": "user", "content": user_prompt + " Now start to give me instructions one by one. Only reply with Instruction and Input."}]
assistant_messages = [
{"role": "system", "content": assistant_prompt},
{"role": "user", "content": assistant_prompt},
]
user_messages = [
{"role": "system", "content": user_prompt},
{
"role": "user",
"content": user_prompt
+ " Now start to give me instructions one by one. Only reply with Instruction and Input.",
},
]
repeat_word_counter = 0
repeat_word_threshold_exceeded = False

for i in range(maxTurns):
repeated_word_current_turn = False

# User turn
if 'claude' in model:
if "claude" in model:
messages = convert_history_to_claude(user_messages)
user_completion = module.completion(
prompt=messages,
stop_sequences=[HUMAN_PROMPT],
model=model,
**hparams
**hparams,
)
user_message = {
"role": "assistant",
"content": user_completion['completion'],
"content": user_completion["completion"],
}
else:
user_completion = module.ChatCompletion.create(
model=model, messages=user_messages, **hparams)
model=model, messages=user_messages, **hparams
)
user_message = user_completion.choices[0].message
user_messages.append(user_message)
assistant_messages.append(
{"role": "user", "content": user_message['content']})
logging.info(
f"User turn {i}: {user_message}")
{"role": "user", "content": user_message["content"]}
)
logging.info(f"User turn {i}: {user_message}")

# Assistant turn
if 'claude' in model:
if "claude" in model:
messages = convert_history_to_claude(assistant_messages)
assistant_completion = module.completion(
prompt=messages,
stop_sequences=[HUMAN_PROMPT],
model=model,
**hparams
**hparams,
)
assistant_message = {
"role": "assistant",
"content": assistant_completion['completion'],
"content": assistant_completion["completion"],
}
else:
assistant_completion = module.ChatCompletion.create(
model=model, messages=assistant_messages, **hparams)
model=model, messages=assistant_messages, **hparams
)
assistant_message = assistant_completion.choices[0].message
assistant_messages.append(assistant_message)
user_messages.append(
{"role": "user", "content": assistant_message['content']})
logging.info(
f"Assistant turn {i}: {assistant_message}")
{"role": "user", "content": assistant_message["content"]}
)
logging.info(f"Assistant turn {i}: {assistant_message}")

yield (user_messages, assistant_messages)

for repeat_word in repeat_word_list:
if repeat_word in assistant_message['content'].lower(
) or repeat_word in user_message['content'].lower():
if (
repeat_word in assistant_message["content"].lower()
or repeat_word in user_message["content"].lower()
):
repeat_word_counter += 1
repeated_word_current_turn = True
logging.info(
f"Repeat word counter = {repeat_word_counter}")
logging.info(f"Repeat word counter = {repeat_word_counter}")
if repeat_word_counter == repeat_word_threshold:
repeat_word_threshold_exceeded = True
break

if not repeated_word_current_turn:
repeat_word_counter = 0

if ("CAMEL_TASK_DONE" in user_messages[-2]['content']) or (i == maxTurns-1) or repeat_word_threshold_exceeded:
if (
("CAMEL_TASK_DONE" in user_messages[-2]["content"])
or (i == maxTurns - 1)
or repeat_word_threshold_exceeded
):
# summary_context = '\n'.join(
# turn['content'] for turn in assistant_messages if turn['role'] == 'assistant')
summary_context = '\n'.join([
f"{turn['role'].capitalize()} ({userRole if turn['role'] == 'user' else assistantRole}): {turn['content']}"
for turn in assistant_messages
if turn['role'] in ['assistant', 'user']
])
summary_context = "\n".join(
[
f"{turn['role'].capitalize()} ({userRole if turn['role'] == 'user' else assistantRole}): {turn['content']}"
for turn in assistant_messages
if turn["role"] in ["assistant", "user"]
]
)

logging.info(f"summary context: {summary_context}")
summary_system_prompt = """You are an experienced solution extracting agent.
Expand All @@ -166,30 +223,34 @@ def camel_agent_generator(userRole, assistantRole, taskPrompt, model, summary_mo
Only use the provided context above and no other sources.
Even if I told you that the task is completed in the context above you should still reply with a complete solution. Never tell me the task is completed or ask for the next request, but instead replay the final solution back to me."""
summary_prompt = (
f"Task:{taskPrompt}\n" + summary_context + summary_closing_prompt)
f"Task:{taskPrompt}\n" + summary_context + summary_closing_prompt
)
summary_messages = [
{"role": "system", "content": summary_system_prompt},
{"role": "user", "content": "Here is the conversation: " + summary_prompt}
{
"role": "user",
"content": "Here is the conversation: " + summary_prompt,
},
]
if 'claude' in model:
if "claude" in model:
summary_completion = module.completion(
prompt=convert_history_to_claude(summary_messages),
stop_sequences=[HUMAN_PROMPT],
model=summary_model,
**hparams
**hparams,
)
message = {
"role": "assistant",
"content": summary_completion['completion'],
"content": summary_completion["completion"],
}
else:
summary_completion = module.ChatCompletion.create(
model=summary_model, messages=summary_messages, temperature=0.2)
model=summary_model, messages=summary_messages, temperature=0.2
)
message = summary_completion.choices[0].message
assistant_messages.append(message)
user_messages.append(
{"role": "user", "content": message['content']})
logging.info(message['content'])
user_messages.append({"role": "user", "content": message["content"]})
logging.info(message["content"])

# End of conversation
yield (user_messages, assistant_messages)
Expand Down
2 changes: 1 addition & 1 deletion pyproject.toml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@ build-backend = "hatchling.build"

[project]
name = "log10-io"
version = "0.0.12"
version = "0.0.13"
authors = []
license = "MIT"
description = "Unified LLM data management"
Expand Down
2 changes: 1 addition & 1 deletion setup.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@

setup(
name="Log10",
version="0.0.12",
version="0.0.13",
description="Log10 LLM data management",
author="Log10 team",
author_email="[email protected]",
Expand Down

0 comments on commit 5312e05

Please sign in to comment.