Spaces:
Paused
Paused
File size: 13,083 Bytes
f647629 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 |
from __future__ import annotations
import json
import os
from typing import Any, Dict, List
import pytest
from wandb_mcp_server.utils import get_rich_logger
from tests.anthropic_test_utils import (
call_anthropic,
check_correctness_tool,
extract_anthropic_tool_use,
get_anthropic_tool_result_message,
)
from wandb_mcp_server.mcp_tools.query_wandbot import (
WANDBOT_TOOL_DESCRIPTION,
query_wandbot_api,
)
from wandb_mcp_server.mcp_tools.tools_utils import generate_anthropic_tool_schema
# -----------------------------------------------------------------------------
# Logging & env guards
# -----------------------------------------------------------------------------
logger = get_rich_logger(__name__)
# -----------------------------------------------------------------------------
# Environment guards
# -----------------------------------------------------------------------------
ANTHROPIC_API_KEY = os.getenv("ANTHROPIC_API_KEY")
WANDBOT_BASE_URL = os.getenv("WANDBOT_TEST_URL", "https://morg--wandbot-api-wandbotapi-serve.modal.run")
if not ANTHROPIC_API_KEY:
pytest.skip(
"ANTHROPIC_API_KEY environment variable not set; skipping Anthropic tests.",
allow_module_level=True,
)
# -----------------------------------------------------------------------------
# Static test context
# -----------------------------------------------------------------------------
MODEL_NAME = "claude-3-7-sonnet-20250219"
CORRECTNESS_MODEL_NAME = "claude-3-5-haiku-20241022"
# -----------------------------------------------------------------------------
# Build tool schema for Anthropic
# -----------------------------------------------------------------------------
available_tools: Dict[str, Dict[str, Any]] = {
"query_wandbot_api": {
"function": query_wandbot_api,
"schema": generate_anthropic_tool_schema(
func=query_wandbot_api, # Pass the function itself
description=WANDBOT_TOOL_DESCRIPTION, # Use the imported description
),
}
}
tools: List[Dict[str, Any]] = [available_tools["query_wandbot_api"]["schema"]]
# -----------------------------------------------------------------------------
# Natural-language queries to test
# -----------------------------------------------------------------------------
test_queries = [
{
"question": "What kinds of scorers does weave support?",
"expected_output": "There are 2 types of scorers in weave, Function-based and Class-based.",
},
# Add more test cases here later
]
# -----------------------------------------------------------------------------
# Tests
# -----------------------------------------------------------------------------
@pytest.mark.parametrize(
"sample",
test_queries,
ids=[f"sample_{i}" for i, _ in enumerate(test_queries)],
)
def test_query_wandbot(sample):
"""End-to-end test: NL question → Anthropic → tool_use → result validation with correctness check."""
query_text = sample["question"]
expected_output = sample[
"expected_output"
] # Get expected output for correctness check
logger.info("\n==============================")
logger.info("QUERY: %s", query_text)
# --- Retry Logic Setup ---
max_retries = 1
last_reasoning = "No correctness check performed yet."
last_is_correct = False
first_call_assistant_response = None # Store the response dict from the first model
tool_result = None # Store the result of executing the tool
tool_use_id = None # Initialize tool_use_id *before* the loop
# Initial messages for the first attempt
messages_first_call = [{"role": "user", "content": query_text}]
for attempt in range(max_retries + 1):
logger.info(f"\n--- Attempt {attempt + 1} / {max_retries + 1} ---")
current_messages = messages_first_call # Start with the base messages
if attempt > 0:
# Retry logic: Add previous assistant response, tool result, and user feedback
retry_messages = []
if first_call_assistant_response:
# 1. Add previous assistant message (contains tool use)
retry_messages.append(
{
"role": first_call_assistant_response.role,
"content": first_call_assistant_response.content,
}
)
# 2. Add the result from executing the tool in the previous attempt
if tool_result is not None and tool_use_id is not None:
tool_result_message = get_anthropic_tool_result_message(
tool_result, tool_use_id
)
retry_messages.append(tool_result_message)
else:
logger.warning(
f"Attempt {attempt + 1}: Cannot add tool result message, tool_result or tool_use_id missing."
)
# 3. Add the user message asking for a retry
retry_user_message_content = f"""
Executing the previous tool call resulted in:
```json
{json.dumps(tool_result, indent=2)}
```
A separate check determined this result was incorrect for the original query.
The reasoning provided was: "{last_reasoning}".
Please re-analyze the original query ("{query_text}") and the result from your previous attempt, then try generating the '{available_tools["query_wandbot_api"]["schema"]["name"]}' tool call again.
"""
retry_messages.append(
{"role": "user", "content": retry_user_message_content}
)
current_messages = (
messages_first_call[:1] + retry_messages
) # Rebuild message list for retry
else:
logger.warning(
"Attempting retry, but no previous assistant response or tool_use_id found."
)
# If retry is needed but we lack context, we probably should just fail or stick with original messages
# For now, let's proceed with original messages, though this might not be ideal.
current_messages = messages_first_call
# --- First Call: Get the query_wandbot_api tool use ---
try:
response = call_anthropic(
model_name=MODEL_NAME,
messages=current_messages, # Use the potentially updated message list
tools=tools,
)
first_call_assistant_response = response # Store for potential *next* retry
except Exception as e:
pytest.fail(f"Attempt {attempt + 1}: Anthropic API call failed: {e}")
try:
# Extract tool_use_id here
_, tool_name, tool_input, tool_use_id = extract_anthropic_tool_use(response)
if tool_use_id is None:
logger.warning(
f"Attempt {attempt + 1}: Model did not return a tool use block."
)
# Decide how to handle this - maybe fail, maybe retry without tool use?
# For now, continue to execution, it might fail gracefully or correctness check will catch it.
except ValueError as e:
logger.error(
f"Attempt {attempt + 1}: Failed to extract tool use from response: {response}"
)
pytest.fail(f"Attempt {attempt + 1}: Could not extract tool use: {e}")
logger.info(f"Attempt {attempt + 1}: Tool emitted by model: {tool_name}")
logger.info(
f"Attempt {attempt + 1}: Tool input: {json.dumps(tool_input, indent=2)}"
)
assert tool_name == "query_wandbot_api", (
f"Attempt {attempt + 1}: Expected 'query_wandbot_api', got '{tool_name}'"
)
assert "question" in tool_input, (
f"Attempt {attempt + 1}: Tool input missing 'question'"
)
# --- Execute the WandBot tool ---
try:
# --- Ensure only expected args based on the *current* function signature are passed ---
# Assuming the function now only takes 'question'
if "question" not in tool_input:
pytest.fail(
f"Attempt {attempt + 1}: Tool input missing required 'question' argument."
)
actual_args = {"question": tool_input["question"]}
tool_result = available_tools[tool_name]["function"](**actual_args)
logger.info(
f"Attempt {attempt + 1}: Tool result: {json.dumps(tool_result, indent=2)}"
) # Log full result
# Basic structure check before correctness check
assert isinstance(tool_result, dict), "Tool result should be a dictionary"
assert isinstance(tool_result.get("answer"), str), (
"'answer' should be a string"
)
assert isinstance(tool_result.get("sources"), list), (
"'sources' should be a list"
)
except Exception as e:
logger.error(
f"Attempt {attempt + 1}: Error executing or validating tool '{tool_name}' with input {actual_args}: {e}",
exc_info=True,
)
pytest.fail(
f"Attempt {attempt + 1}: Tool execution or basic validation failed: {e}"
)
# --- Second Call: Perform Correctness Check ---
logger.info(f"\n--- Starting Correctness Check for Attempt {attempt + 1} ---")
try:
correctness_prompt = f"""
Please evaluate if the provided 'Actual Tool Result' provides a helpful and relevant answer to the 'Original User Query'.
The 'Expected Output Hint' gives guidance on what a good answer should contain.
Use the 'check_correctness_tool' to provide your reasoning and conclusion.
Original User Query:
"{query_text}"
Expected Output:
"{expected_output}"
Actual Tool Result from '{tool_name}':
```json
{json.dumps(tool_result, indent=2)}
```
"""
messages_check_call = [{"role": "user", "content": correctness_prompt}]
correctness_response = call_anthropic(
model_name=CORRECTNESS_MODEL_NAME,
messages=messages_check_call,
check_correctness_tool=check_correctness_tool, # Pass the imported tool schema
)
logger.info(
f"Attempt {attempt + 1}: Correctness check response:\n{correctness_response}\n\n"
)
_, check_tool_name, check_tool_input, _ = extract_anthropic_tool_use(
correctness_response
)
assert check_tool_name == "check_correctness_tool", (
f"Attempt {attempt + 1}: Expected correctness tool, got {check_tool_name}"
)
assert "reasoning" in check_tool_input, (
f"Attempt {attempt + 1}: Correctness tool missing 'reasoning'"
)
assert "is_correct" in check_tool_input, (
f"Attempt {attempt + 1}: Correctness tool missing 'is_correct'"
)
last_reasoning = check_tool_input["reasoning"]
last_is_correct = check_tool_input["is_correct"]
logger.info(
f"Attempt {attempt + 1}: Correctness Reasoning: {last_reasoning}"
)
logger.info(
f"Attempt {attempt + 1}: Is Correct according to LLM: {last_is_correct}"
)
if last_is_correct:
logger.info(
f"--- Correctness check passed on attempt {attempt + 1}. ---"
)
break # Exit the loop successfully
except KeyError as e:
logger.error(
f"Attempt {attempt + 1}: Missing expected key in correctness tool input: {e}"
)
logger.error(
f"Attempt {attempt + 1}: Full input received: {check_tool_input}"
)
last_is_correct = False
last_reasoning = f"Correctness tool response missing key: {e}"
# Continue loop if retries left, fail otherwise handled after loop
except Exception as e:
logger.error(
f"Attempt {attempt + 1}: Error during correctness check for query '{query_text}': {e}",
exc_info=True,
)
last_is_correct = False
last_reasoning = f"Correctness check failed with exception: {e}"
# Continue loop if retries left, fail otherwise handled after loop
# --- After the loop, fail the test if the last attempt wasn't correct ---
if not last_is_correct:
pytest.fail(
f"LLM evaluation failed after {max_retries + 1} attempts. "
f"Final is_correct_flag is `{last_is_correct}`. "
f"Final Reasoning: '{last_reasoning}'"
)
# If we reach here, it means the correctness check passed within the allowed attempts.
logger.info("--- Test passed within allowed attempts. ---")
|