File size: 11,171 Bytes
ead3819
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage, SystemMessage
from langchain.tools import tool

from langchain_community.document_loaders import WikipediaLoader,ArxivLoader

from tavily import TavilyClient

from openai import OpenAI
import base64
import re
import os


from typing import TypedDict, Annotated, Literal
 
from langchain_core.messages import (
    AnyMessage, HumanMessage, AIMessage, ToolMessage, SystemMessage
)

from langgraph.graph.message import add_messages
from langgraph.graph import StateGraph, END



OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")
TAVILY_API_KEY = os.environ.get("TAVILY_API_KEY")


tavily_client = TavilyClient(api_key=TAVILY_API_KEY)
openai_client = OpenAI(api_key=OPENAI_API_KEY)

MAX_STEPS = 15

@tool
def search_wikipedia(query: str, max_docs: int = 3) -> str:
    """Search Wikipedia for general knowledge and return summarized content.
 
    Args:
        query: Topic to search (e.g., 'Artificial Intelligence', 'France history')
        max_docs: Maximum number of Wikipedia pages to retrieve
    """
    loader = WikipediaLoader(query=query, load_max_docs=max_docs)
    docs = loader.load()
    return "\n\n".join(doc.page_content[:3000] for doc in docs)


@tool
def search_arxiv(query: str, max_docs: int = 3) -> str:
    """Search arXiv for scientific papers and return summaries.
 
    Args:
        query: Research topic or keywords (e.g., 'transformer attention')
        max_docs: Maximum number of papers to retrieve
    """
    loader = ArxivLoader(query=query, load_max_docs=max_docs)
    docs = loader.load()
    return "\n\n".join(doc.page_content[:3000] for doc in docs)


@tool
def search_web(query: str, max_results: int = 5) -> str:
    """Search the web for up-to-date information.
 
    Args:
        query: Search query (e.g., 'latest OpenAI model 2025')
        max_results: Number of results to return
    """
    response = tavily_client.search(query=query, max_results=max_results)
    results = [f"{r['title']}\n{r['content']}" for r in response["results"]]
    return "\n\n".join(results)



@tool
def transcribe_audio(file_path: str) -> str:
    """Transcribe an audio file (mp3, wav) into text.
 
    Args:
        file_path: Path to the audio file on disk
    """
    with open(file_path, "rb") as f:
        transcript = openai_client.audio.transcriptions.create(
            model="whisper-1",
            file=f,
        )
    return transcript.text


@tool
def read_image(file_path: str) -> str:
    """Read an image file and return a description via GPT-4o vision.
 
    Args:
        file_path: Path to the image file on disk
    """
    with open(file_path, "rb") as f:
        b64 = base64.b64encode(f.read()).decode("utf-8")
    ext = file_path.rsplit(".", 1)[-1].lower()
    mime = {"jpg": "image/jpeg", "jpeg": "image/jpeg",
            "png": "image/png", "gif": "image/gif",
            "webp": "image/webp"}.get(ext, "image/png")
    response = openai_client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {
                "role": "user",
                "content": [
                    {"type": "image_url",
                     "image_url": {"url": f"data:{mime};base64,{b64}"}},
                    {"type": "text",
                     "text": "Describe this image in detail. Extract any text, data, or key information visible."},
                ],
            }
        ],
        max_tokens=1024,
    )
    return response.choices[0].message.content

@tool
def read_file(file_path: str) -> str:
    """Read a file and return its contents."""
    with open(file_path, "r", encoding="utf-8") as f:
        return f.read()

@tool
def python_repl(code: str) -> str:
    """Execute Python code and return stdout + the value of the last expression.
    Useful for arithmetic, data manipulation, and logic tasks.
 
    Args:
        code: Valid Python code string
    """
    import io, sys, traceback
    stdout_capture = io.StringIO()
    local_vars: dict = {}
    try:
        sys.stdout = stdout_capture
        exec(code, {}, local_vars)           # run all lines
        # try to eval last line as expression
        lines = [l for l in code.strip().splitlines() if l.strip()]
        last_val = ""
        if lines:
            try:
                last_val = repr(eval(lines[-1], {}, local_vars))
            except Exception:
                pass
    except Exception:
        return traceback.format_exc()
    finally:
        sys.stdout = sys.__stdout__
    out = stdout_capture.getvalue()
    return "\n".join(filter(None, [out, last_val])) or "Code executed successfully (no output)."




TOOLS = [
    search_wikipedia,
    search_arxiv,
    search_web,
    transcribe_audio,
    read_image,
    read_file,
    python_repl,
]
 
TOOL_MAP = {t.name: t for t in TOOLS}


SYSTEM_PROMPT = f"""You are a highly capable AI assistant solving tasks from the GAIA benchmark.
 
## Core rules (MUST follow)
1. THINK before acting: decompose the question and plan which tool(s) you need.
2. NEVER call the same tool with the exact same arguments twice.
   If the result was insufficient, use a DIFFERENT query or a DIFFERENT tool.
3. If search_wikipedia returns a biography page instead of a discography/list,
   immediately switch to search_web with a more specific query.
4. For calculations / counting, always use python_repl β€” never guess numbers.
5. Once you have enough information, STOP calling tools and give the final answer.
6. You have at most {MAX_STEPS} tool-call rounds total. Budget them wisely.
 
## Tool selection guide
- General facts / biography  β†’ search_wikipedia (vary query if first try fails)
- Discographies, filmographies, lists β†’ search_web (Wikipedia tool may miss these)
- Current events / live data β†’ search_web
- Scientific papers          β†’ search_arxiv
- Arithmetic / logic         β†’ python_repl
- Provided image file        β†’ read_image
- Provided audio file        β†’ transcribe_audio
- Provided text/csv/json     β†’ read_file
 
## Answer format
End your FINAL response with exactly:
FINAL ANSWER: <your answer>
 
Keep it concise β€” no units unless asked, lists comma-separated.
"""


class AgentState(TypedDict):
    messages:   Annotated[list[AnyMessage], add_messages]
    step_count: int   # counts agent_node invocations


def make_llm(model: str = "gpt-5.4-mini") -> ChatOpenAI:
    return ChatOpenAI(
        model=model,
        temperature=0,
        api_key=OPENAI_API_KEY,
    ).bind_tools(TOOLS)
 
 
llm_with_tools = make_llm()

_step = 0  # console display counter
 
CYAN   = "\033[96m"
GREEN  = "\033[92m"
YELLOW = "\033[93m"
RED    = "\033[91m"
BOLD   = "\033[1m"
RESET  = "\033[0m"
 
 
def _log(label: str, text: str, color: str = RESET) -> None:
    print(f"{color}{'─'*60}{RESET}")
    print(f"{color}[Step {_step}] {label}{RESET}")
    if text.strip():
        print(f"{color}{text.strip()}{RESET}")


def agent_node(state: AgentState) -> AgentState:
    global _step
    _step += 1
    step_count = state.get("step_count", 0) + 1
 
    messages = state["messages"]
 
    # Inject system prompt on first turn
    if not any(isinstance(m, SystemMessage) for m in messages):
        messages = [SystemMessage(content=SYSTEM_PROMPT)] + messages
 
    # Warn model to wrap up when approaching the limit
    if step_count >= MAX_STEPS - 2:
        messages = list(messages) + [HumanMessage(
            content=f"⚠️ You have used {step_count}/{MAX_STEPS} steps. "
                    "Do NOT call any more tools. Synthesise what you have and give FINAL ANSWER now."
        )]
 
    _log("πŸ€– AGENT THINKING …", "", CYAN)
    response = llm_with_tools.invoke(messages)
 
    if response.content:
        _log("πŸ€– AGENT RESPONSE", str(response.content)[:600], CYAN)
 
    if response.tool_calls:
        calls_summary = "\n".join(
            f"  β€’ {tc['name']}({', '.join(f'{k}={repr(v)}' for k, v in tc['args'].items())})"
            for tc in response.tool_calls
        )
        _log("πŸ”§ TOOL CALLS PLANNED", calls_summary, YELLOW)
    else:
        _log("βœ… AGENT FINISHED (no more tool calls)", "", GREEN)
 
    return {"messages": [response], "step_count": step_count}
 


def tool_node(state: AgentState) -> AgentState:
    global _step
    last_msg: AIMessage = state["messages"][-1]
    tool_results: list[ToolMessage] = []
 
    for tc in last_msg.tool_calls:
        _step += 1
        tool_fn = TOOL_MAP.get(tc["name"])
        _log(f"βš™οΈ  RUNNING: {tc['name']}",
             "\n".join(f"  {k}: {repr(v)}" for k, v in tc["args"].items()),
             YELLOW)
 
        if tool_fn is None:
            result = f"ERROR: unknown tool '{tc['name']}'"
            _log("❌ TOOL ERROR", result, RED)
        else:
            try:
                result = tool_fn.invoke(tc["args"])
                preview = str(result)[:500] + ("…" if len(str(result)) > 500 else "")
                _log(f"πŸ“₯ RESULT: {tc['name']}", preview, GREEN)
            except Exception as exc:
                result = f"ERROR calling {tc['name']}: {exc}"
                _log(f"❌ TOOL ERROR: {tc['name']}", result, RED)
 
        tool_results.append(
            ToolMessage(content=str(result), tool_call_id=tc["id"])
        )
 
    return {"messages": tool_results}



def should_continue(state: AgentState) -> Literal["tools", "end"]:
    step_count = state.get("step_count", 0)
 
    if step_count >= MAX_STEPS:
        print(f"{RED}{'─'*60}")
        print(f"β›” MAX_STEPS ({MAX_STEPS}) reached β€” forcing end.{RESET}")
        return "end"
 
    last = state["messages"][-1]
    if isinstance(last, AIMessage) and last.tool_calls:
        return "tools"
 
    return "end"


def build_graph() -> StateGraph:
    g = StateGraph(AgentState)
    g.add_node("agent", agent_node)
    g.add_node("tools", tool_node)
    g.set_entry_point("agent")
    g.add_conditional_edges("agent", should_continue, {"tools": "tools", "end": END})
    g.add_edge("tools", "agent")   # always return to agent after tool use
    return g.compile()
 
 
graph = build_graph()


def run_agent(question: str, file_path: str | None = None) -> str:
    """Run the agent on a GAIA question and return the extracted final answer."""
    global _step
    _step = 0
 
    print(f"\n{BOLD}{'═'*60}{RESET}")
    print(f"{BOLD}❓ QUESTION: {question}{RESET}")
    if file_path:
        print(f"{BOLD}πŸ“Ž FILE: {file_path}{RESET}")
    print(f"{BOLD}{'═'*60}{RESET}\n")
 
    content = question
    if file_path:
        content += f"\n\n[Attached file available at: {file_path}]"
 
    result = graph.invoke({
        "messages": [HumanMessage(content=content)],
        "step_count": 0,
    })
 
    last_msg = result["messages"][-1]
    text = last_msg.content if isinstance(last_msg, AIMessage) else str(last_msg)
 
    match = re.search(r"FINAL ANSWER:\s*(.+)", text, re.IGNORECASE | re.DOTALL)
    answer = match.group(1).strip() if match else text.strip()
 
    print(f"\n{BOLD}{GREEN}{'═'*60}{RESET}")
    print(f"{BOLD}{GREEN}🏁 FINAL ANSWER: {answer}{RESET}")
    print(f"{BOLD}{GREEN}{'═'*60}{RESET}\n")
    return answer