Skip to content

Commit

Permalink
Restructure folder structure, update imports, and add verbose flag to…
Browse files Browse the repository at this point in the history 
… OpenAI in personality_finder_vision.py
  • Loading branch information
@chandralegend
chandralegend committed Sep 15, 2024
1 parent cfdef0e commit 117ac70
Show file tree
Hide file tree
Showing 6 changed files with 277 additions and 81 deletions.
2 changes: 1 addition & 1 deletion examples/personality_finder_vision.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@
from semantix.llms import OpenAI
from semantix.types import Image

llm = OpenAI()
llm = OpenAI(verbose=True)


class Personality(Enum):
Expand Down
112 changes: 112 additions & 0 deletions examples/pii_detector.py
View file

Large diffs are not rendered by default.

70 changes: 44 additions & 26 deletions semantix/inference.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,21 @@
from semantix.types.prompt import Information, OutputHint, Tool, TypeExplanation


def simplify_msgs(messages: list) -> list:
"""Combine the messages with content is string."""
new_messages = [messages[0]]
new_message = {"role": "user", "content": ""}
for i in range(1, len(messages)):
if isinstance(messages[i]["content"], str):
new_message["content"] += messages[i]["content"] + "\n"
else:
new_messages.append(new_message)
new_messages.append(messages[i])
new_message = {"role": "user", "content": ""}
new_messages.append(new_message)
return new_messages


class PromptInfo:
"""Class to represent the prompt information. (According to Meaning-Typed Prompting Technique)."""

Expand All @@ -32,43 +47,31 @@ def __init__(
def get_messages(self, model: BaseLLM) -> list:
"""Get the messages for the prompt."""
messages = [model.system_message]
if self.informations:
messages.append(self.get_info_msg(self.informations, model, "informations"))
messages.append(
{
"role": "user",
"content": f"{model.get_message_desc('action')} {self.action}",
}
)
if self.input_informations:
messages.append(
self.get_info_msg(self.input_informations, model, "input_informations")
)
if self.informations:
messages.append(self.get_info_msg(self.informations, model, "informations"))
if self.context:
messages.append(
{
"role": "user",
"content": f"{model.get_message_desc('context')}\n{self.context}",
}
)
if self.type_explanations:
messages.append(
{
"role": "user",
"content": "\n".join(
[
model.get_message_desc("type_explanations"),
*[str(t) for t in self.type_explanations],
]
),
}
)
messages.append(
{
"role": "user",
"content": f"{model.get_message_desc('return_hint')}\n{self.return_hint}",
}
)
messages.append(
{
"role": "user",
"content": f"{model.get_message_desc('action')}\n{self.action}",
}
)
if self.tools:
messages.append(
{
Expand All @@ -78,6 +81,19 @@ def get_messages(self, model: BaseLLM) -> list:
),
}
)
if self.type_explanations:
messages.append(
{
"role": "user",
"content": "\n".join(
[
model.get_message_desc("type_explanations"),
*[str(t) for t in self.type_explanations],
]
),
}
)
messages = simplify_msgs(messages)
return messages

def get_info_msg(
Expand Down Expand Up @@ -119,7 +135,7 @@ def __init__(

def get_messages(self, model: BaseLLM, output: str) -> list:
"""Get the messages for the extract output prompt."""
messages = []
messages = [model.system_message]
messages.append(
{
"role": "user",
Expand All @@ -129,7 +145,7 @@ def get_messages(self, model: BaseLLM, output: str) -> list:
messages.append(
{
"role": "user",
"content": f"{model.get_message_desc('extract_output_return_hint')}\n{self.return_hint}",
"content": f"{model.get_message_desc('return_hint')}\n{self.return_hint}",
}
)
if self.type_explanations:
Expand All @@ -138,7 +154,7 @@ def get_messages(self, model: BaseLLM, output: str) -> list:
"role": "user",
"content": "\n".join(
[
model.get_message_desc("extract_output_type_explanations"),
model.get_message_desc("type_explanations"),
*[str(t) for t in self.type_explanations],
]
),
Expand All @@ -150,6 +166,7 @@ def get_messages(self, model: BaseLLM, output: str) -> list:
"content": model.EXTRACT_OUTPUT_INSTRUCTION,
}
)
messages = simplify_msgs(messages)
return messages


Expand All @@ -165,7 +182,7 @@ def __init__(

def get_messages(self, model: BaseLLM, output: str, error: str) -> list:
"""Get the messages for the output fix prompt."""
messages = []
messages = [model.system_message]
messages.append(
{
"role": "user",
Expand All @@ -175,7 +192,7 @@ def get_messages(self, model: BaseLLM, output: str, error: str) -> list:
messages.append(
{
"role": "user",
"content": f"{model.get_message_desc('output_fix_return_hint')}\n{self.return_hint}",
"content": f"{model.get_message_desc('return_hint')}\n{self.return_hint}",
}
)
if self.type_explanations:
Expand All @@ -184,7 +201,7 @@ def get_messages(self, model: BaseLLM, output: str, error: str) -> list:
"role": "user",
"content": "\n".join(
[
model.get_message_desc("output_fix_type_explanations"),
model.get_message_desc("type_explanations"),
*[str(t) for t in self.type_explanations],
]
),
Expand All @@ -202,6 +219,7 @@ def get_messages(self, model: BaseLLM, output: str, error: str) -> list:
"content": model.OUTPUT_FIX_INSTRUCTION,
}
)
messages = simplify_msgs(messages)
return messages


Expand Down
130 changes: 94 additions & 36 deletions semantix/llms/base.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,23 +14,33 @@
httpx_logger.setLevel(logging.WARNING)

NORMAL = """
Generate and return the output result(s) only, adhering to the provided Type in the following format
Provide the answer in the desired output type definition. Follow the following template to provide the answer.
[Output] <result>
```output
Only provide the output in this section in the desired output type.
```
""" # noqa E501

REASON = """
Reason and return the output result(s) only, adhering to the provided Type in the following format
[Reasoning] <Reason>
[Output] <Result>
Provide the answer in the desired output type definition. Follow the following template to provide the answer.
```reasoning
Reason Step by Step to achieve the goal.
```
```output
Only provide the output in this section in the desired output type.
```
"""

CHAIN_OF_THOUGHT = """
Generate and return the output result(s) only, adhering to the provided Type in the following format. Perform the operation in a chain of thoughts.(Think Step by Step)
[Chain of Thoughts] <Chain of Thoughts>
[Output] <Result>
Provide the answer in the desired output type definition. Follow the following template to provide the answer.
```chain-of-thoughts
Think Step by Step to achieve the goal.
```
```output
Only provide the output in this section in the desired output type.
```
""" # noqa E501

REACT = """
Expand All @@ -41,38 +51,67 @@
""" # noqa E501

REFLECTION = """
Generate and return the output result(s) only, adhering to the provided Type in the following format. Perform the operation in a chain of thoughts.(Think Step by Step) to get the answer.
Assuming that you are prone to making mistakes, reflect upon the steps you taken and perform the operation again in different procedure with more precision to validate the answer.
Provide the answer in the desired output type definition. Follow the following template to provide the answer.
```chain-of-thoughts
Think Step by Step to achieve the goal.
```
```intermediate-output
Only provide the output in this section in the desired output type.
```
```reflection
Reflect and critique on the thought process and provided answer.
```
```output
Only provide the output in this section in the desired output type.
```
""" # noqa E501

EXTRACT_OUTPUT_INSTRUCTION = """
Above output is not in the desired output format. Extract the output in the desired format. Follow the following template to provide the answer.
[Chain of Thoughts] <Thought>
[Answer] <Answer>
[Reflection] <Reflection>
[Output] <Output>
```output
Only provide the output in this section in the desired output type.
```
""" # noqa E501

OUTPUT_FIX_INSTRUCTION = """
Above output is not in the desired Output Type. Follow the following template to provide the answer.
```debug
Debug the error and fix the output.
```
```output
Only provide the output in this section in the desired output type.
```
""" # noqa E501


class BaseLLM:
"""Base Large Language Model (LLM) class."""

MESSAGE_DESCRIPTIONS = {
"informations": "[Information]",
"input_informations": "[Inputs Information]",
"context": "[Context]",
"type_explanations": "[Type Explanations]",
"return_hint": "[Output Information]",
"action": "[Action]",
"tools": "[Tools]",
"informations": "## Additional Information",
"input_informations": "## Inputs",
"context": "## Context",
"type_explanations": "## Type Definitions",
"return_hint": "## Desired Output Type Definition",
"action": "# **Goal**:",
"tools": "## Tools",
"output_fix_error": "## Error Encountered",
"output_fix_output": "## Previous Output",
"extract_output_output": "## Model Output",
}
SYSTEM_PROMPT = "This is an operation you must perform and return the output values. Neither, the methodology, extra sentences nor the code are not needed. Input/Type formatting: Explanation of the Input (variable_name) (type) = value" # noqa E501
SYSTEM_PROMPT = "You are a python developer with a lot of experience and sticks to the instructions given. You are instructed to not provide code snippets but to follow the instructions to achieve the goal in the desired output." # noqa E501
METHOD_PROMPTS = {
"Normal": NORMAL,
"Reason": REASON,
"Chain-of-Thoughts": CHAIN_OF_THOUGHT,
"ReAct": REACT,
"Reflection": REFLECTION,
}
EXTRACT_OUTPUT_INSTRUCTION = "Above output is not in the desired Output Format/Type. Please provide the output in the desired type. Do not repeat the previously provided output. Important: Do not provide the code or the methodology. Only provide the output in the desired format." # noqa E501
OUTPUT_FIX_INSTRUCTION = "Above output is not in the desired Output Format/Type. Please provide the output in the desired type. Do not repeat the previously provided output. Important: Do not provide the code or the methodology. Only provide the output in the desired format." # noqa E501
EXTRACT_OUTPUT_INSTRUCTION = EXTRACT_OUTPUT_INSTRUCTION
OUTPUT_FIX_INSTRUCTION = OUTPUT_FIX_INSTRUCTION

def __init__(self, verbose: bool = False, max_retries: int = 3) -> None:
"""Initialize the Large Language Model (LLM) client."""
Expand All @@ -99,7 +138,21 @@ def __infer__(self, messages: list, model_params: dict) -> str:
@staticmethod
def _msgs_to_str(messages: list) -> str:
"""Convert the messages to a string."""
return "\n\n".join([str(m["content"]) for m in messages])
x = []
for m in messages:
if not isinstance(m["content"], list):
x.append(m["content"])
else:
x.append(
"".join(
[
i["text"] if i["type"] == "text" else i["type"]
for i in m["content"]
]
)
)
print(x)
return "\n".join(x)

def __call__(self, messages: list, model_params: dict) -> str:
"""Infer a response from the input text."""
Expand All @@ -118,29 +171,31 @@ def resolve_output(
"""Resolve the output string to return the reasoning and output."""
if self.verbose:
logger.info(f"Model Output\n{model_output}")
output_match = re.search(r"\[Output\](.*)", model_output, re.DOTALL)
if not output_match:
outputs = dict(re.findall(r"```(.*?)\n(.*?)```", model_output, re.DOTALL))
if "output" not in outputs:
output = self._extract_output(
model_output,
extract_output_prompt_info,
)
else:
output = output_match.group(1).strip()
return self.to_object(output, output_fix_prompt_info, _globals, _locals)
output = outputs["output"].strip()
obj = self.to_object(output, output_fix_prompt_info, _globals, _locals)
return obj

def _extract_output(
self, model_output: str, extract_output_prompt_info: "ExtractOutputPromptInfo"
) -> str:
"""Extract the output from the meaning out string."""
"""Extract the output from the model output."""
if self.verbose:
logger.info("Extracting output from the meaning out string.")
logger.info("Extracting output from the model output.")
output_extract_messages = extract_output_prompt_info.get_messages(
self, model_output
)
output_extract_output = self.__infer__(output_extract_messages, {})
if self.verbose:
logger.info(f"Extracted Output: {output_extract_output}")
return output_extract_output
outputs = dict(re.findall(r"```(.*?)\n(.*?)```", model_output, re.DOTALL))
return outputs["output"].strip()

def to_object(
self,
Expand All @@ -154,7 +209,7 @@ def to_object(
"""Convert the output string to an object."""
if output_fix_prompt_info.return_hint.type == "str":
return output
if num_retries >= self.max_retries:
if num_retries > self.max_retries:
raise ValueError("Failed to convert output to object. Max tries reached.")
if error:
fixed_output = self._fix_output(output, output_fix_prompt_info, error)
Expand Down Expand Up @@ -186,4 +241,7 @@ def _fix_output(
logger.info(f"Error: {error}, Fixing the output.")
output_fix_messages = output_fix_prompt_info.get_messages(self, output, error)
output_fix_output = self.__infer__(output_fix_messages, {})
return output_fix_output
if self.verbose:
logger.info(f"Fixed Output: {output_fix_output}")
outputs = dict(re.findall(r"```(.*?)\n(.*?)```", output, re.DOTALL))
return outputs["output"].strip()
Loading

0 comments on commit 117ac70

Please sign in to comment.