Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -333,91 +333,91 @@ def should_continue(state: AgentState):
|
|
| 333 |
|
| 334 |
# --- Basic Agent Definition ---
|
| 335 |
class BasicAgent:
|
| 336 |
-
def __init__(self):
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
| 359 |
-
|
| 360 |
-
|
| 361 |
-
|
| 362 |
-
Your goal is to provide the concise, factual answer by strictly following a step-by-step reasoning process.
|
| 363 |
-
|
| 364 |
-
**CRITICAL PROTOCOL: YOU MUST FOLLOW THIS PROCESS**
|
| 365 |
-
|
| 366 |
-
1. **ANALYZE:** Read the question and all messages in the history.
|
| 367 |
-
2. **MANDATORY FIRST STEP:** Your *first* response on *any* new task MUST be a plan in plain text. Do NOT call any tool on your first turn. Write down your logic, what you need, and which tool you *plan* to use next. Failure to provide a plan first will result in incorrect behavior.
|
| 368 |
-
3. **EXECUTE:** After submitting your plan, you will run again. Now, execute the *next* step of your plan by calling the *one* appropriate tool using the correct JSON format.
|
| 369 |
-
4. **ANALYZE TOOL OUTPUT:** You will receive a ToolMessage with the output. You MUST read it carefully.
|
| 370 |
-
5. **REPEAT or FINISH:**
|
| 371 |
-
|
| 372 |
-
|
| 373 |
-
|
| 374 |
-
**RULES:**
|
| 375 |
-
* **NEVER** call a tool on the same turn you write a plan (plain text).
|
| 376 |
-
* **NEVER** use your pre-trained "leaked" knowledge for the final answer. The answer *must* come from a ToolMessage (e.g., from `code_interpreter`'s print() or `search_tool`).
|
| 377 |
-
* **NEVER** answer a logic puzzle from memory. You *must* use `code_interpreter`, ensure it `print()`s the result, analyze that output, and then use that printed result for `final_answer_tool`.
|
| 378 |
-
* **NEVER** call `final_answer_tool` until a tool has explicitly given you the answer in its output.
|
| 379 |
-
* **Error Handling:** If a tool call returns an Error, your next step (Step 1 PLAN) MUST analyze the error message and propose a *different* approach (different tool, different arguments, different logic). Do not retry the exact same failed call.
|
| 380 |
-
|
| 381 |
-
**TOOLS:**
|
| 382 |
-
{tool_descriptions}
|
| 383 |
-
|
| 384 |
-
**TOOL FORMAT (JSON ONLY):**
|
| 385 |
-
Respond ONLY with a single JSON block like this when calling a tool:
|
| 386 |
-
```json
|
| 387 |
-
{{
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
}}
|
| 391 |
-
```
|
| 392 |
-
* Replace `tool_name` with the tool's name. Provide arguments in `tool_input`. Match names/types precisely.
|
| 393 |
-
* Do not add any text before or after the JSON block.
|
| 394 |
-
|
| 395 |
-
Example for final_answer_tool:
|
| 396 |
-
```json
|
| 397 |
-
{{
|
| 398 |
-
|
| 399 |
-
|
| 400 |
-
|
| 401 |
-
|
| 402 |
-
}}
|
| 403 |
-
```
|
| 404 |
-
NOTE: The value for "answer" MUST be a string enclosed in double quotes.
|
| 405 |
-
"""
|
| 406 |
-
|
| 407 |
-
|
| 408 |
-
|
| 409 |
-
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
|
| 413 |
-
|
| 414 |
-
|
| 415 |
-
|
| 416 |
-
|
| 417 |
-
|
| 418 |
-
|
| 419 |
-
|
| 420 |
-
|
| 421 |
|
| 422 |
# --- Agent Node with Robust Parsing Fallback ---
|
| 423 |
def agent_node(state: AgentState):
|
|
|
|
| 333 |
|
| 334 |
# --- Basic Agent Definition ---
|
| 335 |
class BasicAgent:
|
| 336 |
+
def __init__(self):
|
| 337 |
+
print("BasicAgent (LangGraph) initializing...")
|
| 338 |
+
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
|
| 339 |
+
if not GROQ_API_KEY: raise ValueError("GROQ_API_KEY secret is not set!")
|
| 340 |
+
|
| 341 |
+
self.tools = defined_tools
|
| 342 |
+
|
| 343 |
+
# Build tool descriptions separately to avoid f-string backslash issues
|
| 344 |
+
tool_desc_list = []
|
| 345 |
+
for tool in self.tools:
|
| 346 |
+
if tool.name == 'code_interpreter':
|
| 347 |
+
desc = (f"- {tool.name}: Executes Python code. Use for calculations, data manipulation, or logic puzzles.\n"
|
| 348 |
+
f" **CODE INTERPRETER RULES:**\n"
|
| 349 |
+
f" 1. ALWAYS use `print()` for final results.\n"
|
| 350 |
+
f" 2. Write SIMPLE, single-step scripts.\n"
|
| 351 |
+
f" 3. PLAN your next script using plain text output first.\n"
|
| 352 |
+
f" 4. Write reasoning as Python comments (#) before code.\n"
|
| 353 |
+
f" 'pandas' (as pd) is available.")
|
| 354 |
+
else:
|
| 355 |
+
desc = f"- {tool.name}: {tool.description}"
|
| 356 |
+
tool_desc_list.append(desc)
|
| 357 |
+
|
| 358 |
+
tool_descriptions = "\n".join(tool_desc_list)
|
| 359 |
+
|
| 360 |
+
# ==================== SYSTEM PROMPT V4 ====================
|
| 361 |
+
self.system_prompt = f"""You are a highly intelligent and meticulous AI assistant for the GAIA benchmark.
|
| 362 |
+
Your goal is to provide the concise, factual answer by strictly following a step-by-step reasoning process.
|
| 363 |
+
|
| 364 |
+
**CRITICAL PROTOCOL: YOU MUST FOLLOW THIS PROCESS**
|
| 365 |
+
|
| 366 |
+
1. **ANALYZE:** Read the question and all messages in the history.
|
| 367 |
+
2. **MANDATORY FIRST STEP:** Your *first* response on *any* new task MUST be a plan in plain text. Do NOT call any tool on your first turn. Write down your logic, what you need, and which tool you *plan* to use next. Failure to provide a plan first will result in incorrect behavior.
|
| 368 |
+
3. **EXECUTE:** After submitting your plan, you will run again. Now, execute the *next* step of your plan by calling the *one* appropriate tool using the correct JSON format.
|
| 369 |
+
4. **ANALYZE TOOL OUTPUT:** You will receive a ToolMessage with the output. You MUST read it carefully.
|
| 370 |
+
5. **REPEAT or FINISH:**
|
| 371 |
+
* **If more steps are needed:** Go back to step 1 (ANALYZE the new info & PLAN). Write an *updated* plan as plain text (e.g., "The search found X. My next step is to use code_interpreter to process X...").
|
| 372 |
+
* **If the ToolMessage contains the final answer:** You MUST call the `final_answer_tool`. Your answer *must* be derived *only* from the ToolMessage output, not your own knowledge.
|
| 373 |
+
|
| 374 |
+
**RULES:**
|
| 375 |
+
* **NEVER** call a tool on the same turn you write a plan (plain text).
|
| 376 |
+
* **NEVER** use your pre-trained "leaked" knowledge for the final answer. The answer *must* come from a ToolMessage (e.g., from `code_interpreter`'s print() or `search_tool`).
|
| 377 |
+
* **NEVER** answer a logic puzzle from memory. You *must* use `code_interpreter`, ensure it `print()`s the result, analyze that output, and then use that printed result for `final_answer_tool`.
|
| 378 |
+
* **NEVER** call `final_answer_tool` until a tool has explicitly given you the answer in its output.
|
| 379 |
+
* **Error Handling:** If a tool call returns an Error, your next step (Step 1 PLAN) MUST analyze the error message and propose a *different* approach (different tool, different arguments, different logic). Do not retry the exact same failed call.
|
| 380 |
+
|
| 381 |
+
**TOOLS:**
|
| 382 |
+
{tool_descriptions}
|
| 383 |
+
|
| 384 |
+
**TOOL FORMAT (JSON ONLY):**
|
| 385 |
+
Respond ONLY with a single JSON block like this when calling a tool:
|
| 386 |
+
```json
|
| 387 |
+
{{
|
| 388 |
+
"tool": "tool_name",
|
| 389 |
+
"tool_input": {{ "arg_name1": "value1", ... }}
|
| 390 |
+
}}
|
| 391 |
+
```
|
| 392 |
+
* Replace `tool_name` with the tool's name. Provide arguments in `tool_input`. Match names/types precisely.
|
| 393 |
+
* Do not add any text before or after the JSON block.
|
| 394 |
+
|
| 395 |
+
Example for final_answer_tool:
|
| 396 |
+
```json
|
| 397 |
+
{{
|
| 398 |
+
"tool": "final_answer_tool",
|
| 399 |
+
"tool_input": {{
|
| 400 |
+
"answer": "The final answer string here"
|
| 401 |
+
}}
|
| 402 |
+
}}
|
| 403 |
+
```
|
| 404 |
+
NOTE: The value for "answer" MUST be a string enclosed in double quotes.
|
| 405 |
+
"""
|
| 406 |
+
|
| 407 |
+
print("Initializing Groq LLM Endpoint...")
|
| 408 |
+
try:
|
| 409 |
+
chat_llm = ChatGroq(
|
| 410 |
+
temperature=0.01, # Low temperature for factual tasks
|
| 411 |
+
groq_api_key=GROQ_API_KEY,
|
| 412 |
+
model_name="openai/gpt-oss-120b" # <-- Switched Model
|
| 413 |
+
)
|
| 414 |
+
print("✅ Groq LLM Endpoint initialized with openai/gpt-oss-120b.")
|
| 415 |
+
except Exception as e:
|
| 416 |
+
print(f"Error initializing Groq: {e}")
|
| 417 |
+
raise
|
| 418 |
+
|
| 419 |
+
self.llm_with_tools = chat_llm.bind_tools(self.tools)
|
| 420 |
+
print("✅ Tools bound to LLM (using bind_tools).")
|
| 421 |
|
| 422 |
# --- Agent Node with Robust Parsing Fallback ---
|
| 423 |
def agent_node(state: AgentState):
|