OMGeeky/gpt-pilot
1
0
Fork 0
mirror of https://github.com/OMGeeky/gpt-pilot.git synced 2025-12-30 08:03:32 +01:00
Code Issues Packages Projects Releases Wiki Activity

Improved JSON prompting for GPT-4 and recover incomplete JSON responses from Code Llama

Browse Source
This commit is contained in:
Nicholas Albion
2023-09-27 10:53:44 +10:00
parent cf97a1be5e
commit 6dd5a032fa
8 changed files with 108 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
pilot/helpers/agents/test_TechLead.py
View File

@@ -8,9 +8,8 @@ load_dotenv()
from main import get_custom_print
from helpers.agents.TechLead import TechLead, DEVELOPMENT_PLANNING_STEP
from helpers.Project import Project
from test.test_utils import assert_non_empty_string, mock_terminal_size
from test.test_utils import assert_non_empty_string
from test.mock_questionary import MockQuestionary
from utils.function_calling import parse_agent_response
class TestTechLead:
@@ -51,10 +50,8 @@ The development process will include the creation of user stories and tasks, bas
self.project.current_step = DEVELOPMENT_PLANNING_STEP
@pytest.mark.uses_tokens
# @patch('database.database.get_progress_steps', return_value=None)
@patch('helpers.AgentConvo.get_saved_development_step', return_value=None)
@patch('helpers.agents.TechLead.save_progress', return_value=None)
# @patch('os.get_terminal_size', mock_terminal_size)
@patch('helpers.agents.TechLead.get_progress_steps', return_value=None)
def test_create_development_plan(self, mock_get_saved_step, mock_save_progress, mock_get_progress_steps):
self.techLead = TechLead(self.project)

2
pilot/prompts/development/plan.prompt
View File

@@ -40,4 +40,4 @@ Each task needs to be related only to the development of this app and nothing el
For each task, there must be a way for human developer to check if the task is done or not. Write how should the developer check if the task is done.
Now, based on the app's description, user stories and user tasks, and the technologies that you need to use, think task by task and write up the entire plan for the development. Start from the project setup and specify each task until the moment when the entire app should be fully working. For each task, write a description and a user-review goal.
Now, based on the app's description, user stories and user tasks, and the technologies that you need to use, think task by task and create the entire development plan. Start from the project setup and specify each task until the moment when the entire app should be fully working. For each task, write a description and a user-review goal.

7
pilot/prompts/utils/incomplete_json.prompt Normal file
View File

@@ -0,0 +1,7 @@
[INST]I received an incomplete JSON response. Please provide the remainder of the JSON object. I will append your entire response to the incomplete JSON data below so it is important that you must not include any of the data already received or any text that does not complete the JSON data.
A response which starts with "Here is the remainder of the JSON object" would be an example of an invalid response, a preamble must NOT be included.
Note that because the JSON data I have already received is an incomplete JSON object, you will need to include the opening and closing curly braces in your response, but rather continue off from EXACTLY where the received JSON ends.
JSON received:
[/INST]
{{ received_json }}

12
pilot/utils/function_calling.py
View File

@@ -70,8 +70,7 @@ def parse_agent_response(response, function_calls: FunctionCallSet | None):
"""
if function_calls:
text = re.sub(r'^.*```json\s*', '', response['text'], flags=re.DOTALL)
text = text.strip('` \n')
text = response['text']
values = list(json.loads(text).values())
if len(values) == 1:
return values[0]
@@ -140,7 +139,7 @@ class JsonPrompter:
return "\n".join(
self.function_descriptions(functions, function_to_call)
+ [
"The response MUST be a JSON object matching this schema:",
"Here is the schema for the expected JSON object:",
"```json",
self.function_parameters(functions, function_to_call),
"```",
@@ -194,7 +193,7 @@ class JsonPrompter:
system = (
"Help choose the appropriate function to call to answer the user's question."
if function_to_call is None
else f"Define the arguments for {function_to_call} to answer the user's question."
else f"Please provide a JSON object that defines the arguments for the `{function_to_call}` function to answer the user's question."
) + "\nThe response must contain ONLY the JSON object, with NO additional text or explanation."
data = (
@@ -202,11 +201,6 @@ class JsonPrompter:
if function_to_call
else self.functions_summary(functions)
)
response_start = (
f"Here are the arguments for the `{function_to_call}` function: ```json\n"
if function_to_call
else "Here's the function the user should call: "
)
if self.is_instruct:
return f"[INST] <<SYS>>\n{system}\n\n{data}\n<</SYS>>\n\n{prompt} [/INST]"

41
pilot/utils/llm_connection.py
View File

@@ -13,7 +13,7 @@ from typing import List
from const.llm import MIN_TOKENS_FOR_GPT_RESPONSE, MAX_GPT_MODEL_TOKENS
from logger.logger import logger
from helpers.exceptions.TokenLimitError import TokenLimitError
from utils.utils import fix_json
from utils.utils import fix_json, get_prompt
from utils.function_calling import add_function_calls_to_request, FunctionCallSet, FunctionType
@@ -148,6 +148,11 @@ def retry_on_exception(func):
err_str = str(e)
# If the specific error "context_length_exceeded" is present, simply return without retry
if isinstance(e, json.JSONDecodeError):
# codellama-34b-instruct seems to send incomplete JSON responses
if e.msg == 'Expecting value':
args[0]['function_buffer'] = e.doc
continue
if "context_length_exceeded" in err_str:
raise TokenLimitError(get_tokens_in_messages_from_openai_error(err_str), MAX_GPT_MODEL_TOKENS)
if "rate_limit_exceeded" in err_str:
@@ -187,14 +192,20 @@ def stream_gpt_completion(data, req_type):
# TODO add type dynamically - this isn't working when connected to the external process
terminal_width = 50 # os.get_terminal_size().columns
lines_printed = 2
gpt_response = ''
buffer = '' # A buffer to accumulate incoming data
expecting_json = False
expecting_json = None
received_json = False
if 'functions' in data:
expecting_json = data['functions']
if 'function_buffer' in data:
incomplete_json = get_prompt('utils/incomplete_json.prompt', {'received_json': data['function_buffer']})
data['messages'].append({'role': 'user', 'content': incomplete_json})
gpt_response = data['function_buffer']
received_json = True
# Don't send the `functions` parameter to Open AI, but don't remove it from `data` in case we need to retry
data = {key: value for key, value in data.items() if key != "functions"}
data = {key: value for key, value in data.items() if not key.startswith('function')}
def return_result(result_data, lines_printed):
if buffer:
@@ -251,7 +262,6 @@ def stream_gpt_completion(data, req_type):
logger.debug(f'problem with request: {response.text}')
raise Exception(f"API responded with status code: {response.status_code}. Response text: {response.text}")
gpt_response = ''
# function_calls = {'name': '', 'arguments': ''}
for line in response.iter_lines():
@@ -283,11 +293,9 @@ def stream_gpt_completion(data, req_type):
# return return_result({'function_calls': function_calls}, lines_printed)
json_line = choice['delta']
# TODO: token healing? https://github.com/1rgs/jsonformer-claude
# ...Is this what local_llm_function_calling.constrainer is for?
except json.JSONDecodeError:
logger.error(f'Unable to decode line: {line}')
except json.JSONDecodeError as e:
logger.error(f'Unable to decode line: {line} {e.msg}')
continue # skip to the next line
# handle the streaming response
@@ -306,16 +314,9 @@ def stream_gpt_completion(data, req_type):
buffer += content # accumulate the data
# If you detect a natural breakpoint (e.g., line break or end of a response object), print & count:
if buffer.endswith("\n"):
if buffer.endswith('\n'):
if expecting_json and not received_json:
received_json = assert_json_response(buffer, lines_printed > 2)
if received_json:
gpt_response = ""
# if not received_json:
# # Don't append to gpt_response, but increment lines_printed
# lines_printed += 1
# buffer = ""
# continue
# or some other condition that denotes a breakpoint
lines_printed += count_lines_based_on_width(buffer, terminal_width)
@@ -333,6 +334,7 @@ def stream_gpt_completion(data, req_type):
logger.info(f'Response message: {gpt_response}')
if expecting_json:
gpt_response = clean_json_response(gpt_response)
assert_json_schema(gpt_response, expecting_json)
new_code = postprocessing(gpt_response, req_type) # TODO add type dynamically
@@ -348,12 +350,17 @@ def assert_json_response(response: str, or_fail=True) -> bool:
return False
def clean_json_response(response: str) -> str:
response = re.sub(r'^.*```json\s*', '', response, flags=re.DOTALL)
return response.strip('` \n')
def assert_json_schema(response: str, functions: list[FunctionType]) -> True:
for function in functions:
schema = function['parameters']
parsed = json.loads(response)
validate(parsed, schema)
return True
return True
def postprocessing(gpt_response, req_type):

1
pilot/utils/test_files.py
View File

@@ -1,4 +1,3 @@
import pytest
from .files import setup_workspace

25
pilot/utils/test_function_calling.py
View File

@@ -1,4 +1,5 @@
from const.function_calls import ARCHITECTURE, DEV_STEPS
from const.function_calls import ARCHITECTURE
from utils.llm_connection import clean_json_response
from .function_calling import parse_agent_response, JsonPrompter
@@ -30,6 +31,7 @@ class TestFunctionCalling:
function_calls = {'definitions': [], 'functions': {}}
# When
response['text'] = clean_json_response(response['text'])
response = parse_agent_response(response, function_calls)
# Then
@@ -41,6 +43,7 @@ class TestFunctionCalling:
function_calls = {'definitions': [], 'functions': {}}
# When
response['text'] = clean_json_response(response['text'])
response = parse_agent_response(response, function_calls)
# Then
@@ -68,7 +71,7 @@ def test_json_prompter():
# Then
assert prompt == '''Help choose the appropriate function to call to answer the user's question.
The response should contain only the JSON object, with no additional text or explanation.
The response must contain ONLY the JSON object, with NO additional text or explanation.
Available functions:
- process_technologies - Print the list of technologies that are created.
@@ -86,7 +89,7 @@ def test_llama_json_prompter():
# Then
assert prompt == '''[INST] <<SYS>>
Help choose the appropriate function to call to answer the user's question.
The response should contain only the JSON object, with no additional text or explanation.
The response must contain ONLY the JSON object, with NO additional text or explanation.
Available functions:
- process_technologies - Print the list of technologies that are created.
@@ -103,11 +106,11 @@ def test_json_prompter_named():
prompt = prompter.prompt('Create a web-based chat app', ARCHITECTURE['definitions'], 'process_technologies')
# Then
assert prompt == '''Define the arguments for process_technologies to answer the user's question.
The response should contain only the JSON object, with no additional text or explanation.
assert prompt == '''Please provide a JSON object that defines the arguments for the `process_technologies` function to answer the user's question.
The response must contain ONLY the JSON object, with NO additional text or explanation.
Print the list of technologies that are created.
The response should be a JSON object matching this schema:
# process_technologies: Print the list of technologies that are created.
Here is the schema for the expected JSON object:
```json
{
"technologies": {
@@ -133,11 +136,11 @@ def test_llama_json_prompter_named():
# Then
assert prompt == '''[INST] <<SYS>>
Define the arguments for process_technologies to answer the user's question.
The response should contain only the JSON object, with no additional text or explanation.
Please provide a JSON object that defines the arguments for the `process_technologies` function to answer the user's question.
The response must contain ONLY the JSON object, with NO additional text or explanation.
Print the list of technologies that are created.
The response should be a JSON object matching this schema:
# process_technologies: Print the list of technologies that are created.
Here is the schema for the expected JSON object:
```json
{
"technologies": {

82
pilot/utils/test_llm_connection.py
View File

@@ -2,6 +2,7 @@ import builtins
from json import JSONDecodeError
import pytest
from unittest.mock import patch, Mock
from dotenv import load_dotenv
from jsonschema import ValidationError
@@ -12,7 +13,8 @@ from helpers.agents.Architect import Architect
from helpers.agents.TechLead import TechLead
from utils.function_calling import parse_agent_response, FunctionType
from test.test_utils import assert_non_empty_string
from .llm_connection import create_gpt_chat_completion, assert_json_response, assert_json_schema
from test.mock_questionary import MockQuestionary
from utils.llm_connection import create_gpt_chat_completion, stream_gpt_completion, assert_json_response, assert_json_schema
from main import get_custom_print
load_dotenv()
@@ -98,14 +100,42 @@ class TestLlmConnection:
def setup_method(self):
builtins.print, ipc_client_instance = get_custom_print({})
@patch('utils.llm_connection.requests.post')
def test_stream_gpt_completion(self, mock_post):
# Given streaming JSON response
deltas = ['{', '\\n',
' \\"foo\\": \\"bar\\",', '\\n',
' \\"prompt\\": \\"Hello\\",', '\\n',
' \\"choices\\": []', '\\n',
'}']
lines_to_yield = [
('{"id": "gen-123", "choices": [{"index": 0, "delta": {"role": "assistant", "content": "' + delta + '"}}]}')
.encode('utf-8')
for delta in deltas
]
lines_to_yield.insert(1, b': OPENROUTER PROCESSING') # Simulate OpenRoute keep-alive pings
mock_response = Mock()
mock_response.status_code = 200
mock_response.iter_lines.return_value = lines_to_yield
mock_post.return_value = mock_response
# When
with patch('utils.llm_connection.requests.post', return_value=mock_response):
response = stream_gpt_completion({}, '')
# Then
assert response == {'text': '{\n "foo": "bar",\n "prompt": "Hello",\n "choices": []\n}'}
@pytest.mark.uses_tokens
@pytest.mark.parametrize("endpoint, model", [
("OPENAI", "gpt-4"), # role: system
("OPENROUTER", "openai/gpt-3.5-turbo"), # role: user
("OPENROUTER", "meta-llama/codellama-34b-instruct"), # rule: user, is_llama
("OPENROUTER", "google/palm-2-chat-bison"), # role: user/system
("OPENROUTER", "google/palm-2-codechat-bison"),
("OPENROUTER", "anthropic/claude-2"), # role: user, is_llama
@pytest.mark.parametrize('endpoint, model', [
('OPENAI', 'gpt-4'), # role: system
('OPENROUTER', 'openai/gpt-3.5-turbo'), # role: user
('OPENROUTER', 'meta-llama/codellama-34b-instruct'), # rule: user, is_llama
('OPENROUTER', 'google/palm-2-chat-bison'), # role: user/system
('OPENROUTER', 'google/palm-2-codechat-bison'),
('OPENROUTER', 'anthropic/claude-2'), # role: user, is_llama
])
def test_chat_completion_Architect(self, endpoint, model, monkeypatch):
# Given
@@ -154,13 +184,13 @@ solution-oriented decision-making in areas where precise instructions were not p
assert 'Node.js' in response
@pytest.mark.uses_tokens
@pytest.mark.parametrize("endpoint, model", [
("OPENAI", "gpt-4"), # role: system
("OPENROUTER", "openai/gpt-3.5-turbo"), # role: user
("OPENROUTER", "meta-llama/codellama-34b-instruct"), # rule: user, is_llama
("OPENROUTER", "google/palm-2-chat-bison"), # role: user/system
("OPENROUTER", "google/palm-2-codechat-bison"),
("OPENROUTER", "anthropic/claude-2"), # role: user, is_llama
@pytest.mark.parametrize('endpoint, model', [
('OPENAI', 'gpt-4'),
('OPENROUTER', 'openai/gpt-3.5-turbo'),
('OPENROUTER', 'meta-llama/codellama-34b-instruct'),
('OPENROUTER', 'google/palm-2-chat-bison'),
('OPENROUTER', 'google/palm-2-codechat-bison'),
('OPENROUTER', 'anthropic/claude-2'),
])
def test_chat_completion_TechLead(self, endpoint, model, monkeypatch):
# Given
@@ -191,18 +221,22 @@ The development process will include the creation of user stories and tasks, bas
})
function_calls = DEVELOPMENT_PLAN
# Retry on bad LLM responses
mock_questionary = MockQuestionary(['', '', 'no'])
# When
response = create_gpt_chat_completion(convo.messages, '', function_calls=function_calls)
with patch('utils.llm_connection.questionary', mock_questionary):
response = create_gpt_chat_completion(convo.messages, '', function_calls=function_calls)
# Then
assert convo.messages[0]['content'].startswith('You are a tech lead in a software development agency')
assert convo.messages[1]['content'].startswith('You are working in a software development agency and a project manager and software architect approach you')
# Then
assert convo.messages[0]['content'].startswith('You are a tech lead in a software development agency')
assert convo.messages[1]['content'].startswith('You are working in a software development agency and a project manager and software architect approach you')
assert response is not None
response = parse_agent_response(response, function_calls)
assert_non_empty_string(response[0]['description'])
assert_non_empty_string(response[0]['programmatic_goal'])
assert_non_empty_string(response[0]['user_review_goal'])
assert response is not None
response = parse_agent_response(response, function_calls)
assert_non_empty_string(response[0]['description'])
assert_non_empty_string(response[0]['programmatic_goal'])
assert_non_empty_string(response[0]['user_review_goal'])
# def test_break_down_development_task(self):