forked from ScottLogic/InferLLM
-
Notifications
You must be signed in to change notification settings - Fork 1
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Co-authored-by: Emma Pearce <[email protected]> Co-authored-by: Maxwell Nyamunda <[email protected]>
- Loading branch information
1 parent
1c3321d
commit 66a6ae4
Showing
9 changed files
with
186 additions
and
5 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Empty file.
38 changes: 38 additions & 0 deletions
38
backend/tests/BDD/features/Correctness/Accuracy_Factual_Correctness.feature
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,38 @@ | ||
@database_agent @ESG | ||
Scenario Outline: When a user asks InferESG for information about their transaction history | ||
Given a prompt to InferESG | ||
When I get the response | ||
Then the response to this '<prompt>' should match the '<expected_response>' | ||
Examples: | ||
|prompt |expected_response | | ||
|Check the database and tell me the average ESG score (Environmental) for the WhiteRock ETF fund |The average ESG score (Environmental) for the WhiteRock ETF fund is approximately 69.67| | ||
|Check the database and tell me the fund with the highest ESG social score |Dynamic Industries with a score of 91| | ||
|Check the database and tell me the fund with the lowest Governance ESG score |Dynamic Industries, which has a score of 60| | ||
# |Check the database and tell me the fund with the lowest ESG score |Dynamic Industries with a score of 50| | ||
# |Check the database and tell me the largest fund |The largest fund is the Global Energy Fund, which has a size of 1,500| | ||
# |Check the database and tell me which funds contain Shell |Funds containing Shell are European Growth Fund, Global Energy Fund, Silverman Global ETF and WhiteRock ETF| | ||
|
||
|
||
@web_agent | ||
Scenario Outline: When a user asks InferESG generic questions | ||
Given a prompt to InferESG | ||
When I get the response | ||
Then the response to this '<prompt>' should match the '<expected_response>' | ||
Examples: | ||
|prompt |expected_response | | ||
|What is the capital of France? |Paris | | ||
|What is the capital of Zimbabwe? |Harare | | ||
|What is the capital of Spain? |Madrid | | ||
|What is the capital of China? |Beijing | | ||
|What is the capital of United Kingdom? |London | | ||
|What is the capital of Sweden? |Stockholm | | ||
|
||
@confidence | ||
Scenario Outline: Check Response's confidence | ||
Given a prompt to InferESG | ||
When I get the response | ||
Then the response to this '<prompt>' should give a confident answer | ||
Examples: | ||
|prompt | | ||
|What is the capital of France? | | ||
|
Empty file.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,85 @@ | ||
from pytest_bdd import given, when, then, parsers, scenarios | ||
import pytest | ||
import logging | ||
from tests.BDD.test_utilities import ( | ||
send_prompt, | ||
app_healthcheck, | ||
correctness_evaluator, | ||
healthy_response, | ||
check_response_confidence, | ||
) | ||
from decimal import Decimal | ||
import decimal | ||
|
||
logger = logging.getLogger(__name__) | ||
|
||
scenarios("../features/Correctness/Accuracy_Factual_Correctness.feature") | ||
|
||
|
||
@pytest.fixture | ||
def context(): | ||
return {} | ||
|
||
|
||
@given(parsers.parse("a prompt to InferESG")) | ||
def prepare_prompt(context): | ||
healthcheck_response = app_healthcheck() | ||
assert healthcheck_response.status_code == 200 | ||
assert healthcheck_response.json() == healthy_response | ||
context["health_check_passed"] = True | ||
|
||
|
||
@when(parsers.parse("I get the response")) | ||
def get_response(context): | ||
assert context.get("health_check_passed", False) | ||
|
||
|
||
@then(parsers.parse("the response to this '{prompt}' should match the '{expected_response}'")) | ||
def check_response_includes_expected_response(context, prompt, expected_response): | ||
response = send_prompt(prompt) | ||
actual_response = response.json() | ||
|
||
try: | ||
expected_value = Decimal(str(expected_response).strip()) | ||
actual_value = Decimal(str(actual_response).strip()) | ||
|
||
tolerance = Decimal("0.01") | ||
is_equal = abs(expected_value - actual_value) <= tolerance | ||
|
||
if not is_equal: | ||
pytest.fail(f"\nNumeric values don't match!\n" f"Expected: {expected_value}\n" f"Actual: {actual_value}") | ||
|
||
except (ValueError, decimal.InvalidOperation): | ||
expected_str = str(expected_response).strip() | ||
actual_str = str(actual_response).strip() | ||
|
||
logger.info(f"Expected : {expected_str} \nActual: {actual_str}") | ||
|
||
if actual_str.find(expected_str) == -1: | ||
result = correctness_evaluator.evaluate_strings( | ||
input=prompt, | ||
prediction=expected_str, | ||
reference=actual_str, | ||
) | ||
|
||
if result["value"] == "N": | ||
logger.error( | ||
f"\nTest failed!\n" | ||
f"Expected: {expected_str}\n" | ||
f"Actual: {actual_str}\n" | ||
f"Reasoning: {result.get('reasoning', 'No reasoning provided')}" | ||
) | ||
|
||
assert result["value"] == "Y", ( | ||
f"\nTest failed!\n" | ||
f"Expected: {expected_str}\n" | ||
f"Actual: {actual_str}\n" | ||
f"Reasoning: {result.get('reasoning', 'No reasoning provided')}" | ||
) | ||
|
||
|
||
@then(parsers.parse("the response to this '{prompt}' should give a confident answer")) | ||
def check_bot_response_confidence(prompt): | ||
response = send_prompt(prompt) | ||
result = check_response_confidence(prompt, response.json()) | ||
assert result["score"] == 1, "The bot response is not confident enough. \nReasoning: " + result["reasoning"] |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,50 @@ | ||
from src.api import app | ||
from src.utils import Config | ||
from fastapi.testclient import TestClient | ||
from langchain.evaluation import EvaluatorType, StringEvaluator, load_evaluator | ||
from langchain_openai.chat_models import ChatOpenAI | ||
|
||
START_ENDPOINT_URL = "/chat?utterance={utterance}" | ||
CONVERSATION_ENDPOINT_URL = "/chat?utterance={utterance}" | ||
HEALTHCHECK_ENDPOINT_URL = "/health" | ||
health_prefix = "InferESG healthcheck: " | ||
healthy_response = health_prefix + "backend is healthy. Neo4J is healthy." | ||
|
||
client = TestClient(app) | ||
config = Config() | ||
|
||
|
||
def app_healthcheck(): | ||
healthcheck_response = client.get(HEALTHCHECK_ENDPOINT_URL) | ||
return healthcheck_response | ||
|
||
|
||
def send_prompt(prompt: str): | ||
start_response = client.get(START_ENDPOINT_URL.format(utterance=prompt)) | ||
return start_response | ||
|
||
# Evaluators | ||
# Evaluation LLM | ||
llm = ChatOpenAI(api_key=config.openai_key, model="gpt-4o-mini", temperature=0, max_retries=2) # type: ignore | ||
|
||
correctness_evaluator: StringEvaluator = load_evaluator( # type: ignore | ||
EvaluatorType.LABELED_CRITERIA, criteria="correctness", llm=llm | ||
) | ||
|
||
confidence_criterion = { | ||
"confidence": "Does the bot seem confident that it replied to the question and gave the correct answer?" | ||
} | ||
|
||
confidence_evaluator: StringEvaluator = load_evaluator( # type: ignore | ||
EvaluatorType.CRITERIA, criteria=confidence_criterion, llm=llm | ||
) | ||
|
||
|
||
def check_response_confidence(prompt: str, bot_response: str) -> dict[str, str]: | ||
""" | ||
Uses an LLM to check the confidence of the bot's response.\n | ||
Returns a dictionary with the binary score (pass = 1, fail = 0) and reasoning (text format).""" | ||
return confidence_evaluator.evaluate_strings( | ||
input=prompt, | ||
prediction=bot_response, | ||
) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters