dawid-lorek commited on
Commit
d092bd1
Β·
verified Β·
1 Parent(s): 48d9442

Update agent.py

Browse files
Files changed (1) hide show
  1. agent.py +55 -29
agent.py CHANGED
@@ -1,15 +1,14 @@
1
- # agent.py β€” final version without circular imports
2
 
3
  import os
4
  import asyncio
5
  from llama_index.llms.openai import OpenAI
6
- from llama_index.core.agent.react.base import ReActAgent
7
  from llama_index.core.tools import FunctionTool
8
 
9
- from langchain_community.tools.wikipedia.tool import WikipediaQueryRun
10
  from langchain_community.utilities.wikipedia import WikipediaAPIWrapper
11
- from langchain_experimental.tools.python.tool import PythonREPLTool
12
  from langchain_community.document_loaders import YoutubeLoader
 
13
 
14
  import whisper
15
  import openpyxl
@@ -20,34 +19,51 @@ if os.getenv("OPENAI_API_KEY"):
20
  else:
21
  print("⚠️ Missing OPENAI_API_KEY – LLM may fail")
22
 
23
- # Tools definitions
24
- api_wrapper = WikipediaAPIWrapper(top_k_results=1, doc_content_chars_max=1000)
 
 
 
 
25
 
26
- def wikipedia_search(query: str) -> str:
27
- return WikipediaQueryRun(api_wrapper=api_wrapper).run({"query": query})
 
 
 
 
 
 
 
 
 
28
 
29
- def run_python_with_output(code: str) -> str:
30
- if "print(" not in code:
31
- code = f"print({code})"
32
- return PythonREPLTool().run(code)
33
 
34
  def get_youtube_transcript(url: str) -> str:
 
35
  try:
36
  loader = YoutubeLoader.from_youtube_url(url, add_video_info=False)
37
  docs = loader.load()
38
  return " ".join(d.page_content for d in docs)
39
  except Exception as e:
40
- return "[YOUTUBE ERROR] " + str(e)
 
 
41
 
42
  def transcribe_audio(file_path: str) -> str:
 
43
  try:
44
  model = whisper.load_model("base")
45
  res = model.transcribe(file_path)
46
  return res["text"]
47
  except Exception as e:
48
- return "[AUDIO ERROR] " + str(e)
 
 
49
 
50
  def extract_excel_total_food_sales(file_path: str) -> str:
 
51
  try:
52
  wb = openpyxl.load_workbook(file_path)
53
  sheet = wb.active
@@ -57,31 +73,41 @@ def extract_excel_total_food_sales(file_path: str) -> str:
57
  total += float(amount or 0)
58
  return f"${total:.2f}"
59
  except Exception as e:
60
- return "[EXCEL ERROR] " + str(e)
61
 
62
- # Assemble tools
63
  TOOLS = [
64
- FunctionTool.from_defaults(wikipedia_search),
65
- FunctionTool.from_defaults(run_python_with_output),
66
- FunctionTool.from_defaults(get_youtube_transcript),
67
- FunctionTool.from_defaults(transcribe_audio),
68
- FunctionTool.from_defaults(extract_excel_total_food_sales),
69
  ]
70
 
71
- # LLM and Agent
72
  llm = OpenAI(model="gpt-4")
73
- agent = ReActAgent.from_tools(
 
74
  tools=TOOLS,
75
  llm=llm,
76
  verbose=True,
77
  system_prompt="""
78
- You are an expert AI assistant on the GAIA benchmark.
 
 
 
 
 
 
 
79
 
80
- Use available tools (Wikipedia, Python, YouTube transcript, audio, Excel).
81
- Output ONLY the final answer. No reasoning or commentary.
82
- Format exactly as requested (list, number, name, chess move, currency).
83
- If tool fails, output "Tool not available".
84
- """,
 
 
85
  )
86
 
87
  def answer_question_sync(question: str) -> str:
 
1
+ # agent.py β€” GAIA-ready with FunctionCallingAgent and improved tools
2
 
3
  import os
4
  import asyncio
5
  from llama_index.llms.openai import OpenAI
6
+ from llama_index.core.agent import FunctionCallingAgent
7
  from llama_index.core.tools import FunctionTool
8
 
 
9
  from langchain_community.utilities.wikipedia import WikipediaAPIWrapper
 
10
  from langchain_community.document_loaders import YoutubeLoader
11
+ from langchain_experimental.tools.python.tool import PythonREPLTool
12
 
13
  import whisper
14
  import openpyxl
 
19
  else:
20
  print("⚠️ Missing OPENAI_API_KEY – LLM may fail")
21
 
22
+ # Tool 1 β€” Wikipedia
23
+ wiki_api = WikipediaAPIWrapper(top_k_results=1, doc_content_chars_max=1000)
24
+
25
+ def search_wikipedia(query: str) -> str:
26
+ """Search Wikipedia for a given query and return relevant summary."""
27
+ return wiki_api.run(query)
28
 
29
+ # Tool 2 β€” Python with output
30
+ python_tool = PythonREPLTool()
31
+
32
+ def run_python_code(code: str) -> str:
33
+ """Run Python code and return printed result."""
34
+ try:
35
+ if "print(" not in code:
36
+ code = f"print({code})"
37
+ return python_tool.run(code)
38
+ except Exception as e:
39
+ return f"[PYTHON ERROR] {e}"
40
 
41
+ # Tool 3 β€” YouTube transcript
 
 
 
42
 
43
  def get_youtube_transcript(url: str) -> str:
44
+ """Get transcript from YouTube video."""
45
  try:
46
  loader = YoutubeLoader.from_youtube_url(url, add_video_info=False)
47
  docs = loader.load()
48
  return " ".join(d.page_content for d in docs)
49
  except Exception as e:
50
+ return f"[YOUTUBE ERROR] {e}"
51
+
52
+ # Tool 4 β€” Whisper transcription
53
 
54
  def transcribe_audio(file_path: str) -> str:
55
+ """Transcribe an MP3 file to text using Whisper."""
56
  try:
57
  model = whisper.load_model("base")
58
  res = model.transcribe(file_path)
59
  return res["text"]
60
  except Exception as e:
61
+ return f"[AUDIO ERROR] {e}"
62
+
63
+ # Tool 5 β€” Excel parser
64
 
65
  def extract_excel_total_food_sales(file_path: str) -> str:
66
+ """Sum sales from Excel where category is 'food'."""
67
  try:
68
  wb = openpyxl.load_workbook(file_path)
69
  sheet = wb.active
 
73
  total += float(amount or 0)
74
  return f"${total:.2f}"
75
  except Exception as e:
76
+ return f"[EXCEL ERROR] {e}"
77
 
78
+ # Assemble tools with proper descriptions
79
  TOOLS = [
80
+ FunctionTool.from_defaults(search_wikipedia, name="search_wikipedia", description="Search Wikipedia for facts and lists."),
81
+ FunctionTool.from_defaults(run_python_code, name="run_python", description="Run Python code for logic, math, or set processing."),
82
+ FunctionTool.from_defaults(get_youtube_transcript, name="get_youtube_transcript", description="Fetch transcript from YouTube video by URL."),
83
+ FunctionTool.from_defaults(transcribe_audio, name="transcribe_audio", description="Transcribe MP3 audio file using Whisper."),
84
+ FunctionTool.from_defaults(extract_excel_total_food_sales, name="extract_excel_total_food_sales", description="Sum total sales from Excel where category is 'food'.")
85
  ]
86
 
87
+ # Create agent with improved system prompt
88
  llm = OpenAI(model="gpt-4")
89
+
90
+ agent = FunctionCallingAgent.from_tools(
91
  tools=TOOLS,
92
  llm=llm,
93
  verbose=True,
94
  system_prompt="""
95
+ You are a highly capable AI agent taking the GAIA benchmark test.
96
+
97
+ You have access to the following tools:
98
+ - Wikipedia search for factual lookups
99
+ - Python runner for math, logic, or text analysis
100
+ - YouTube transcript fetcher (via URL)
101
+ - Audio transcriber (Whisper, MP3)
102
+ - Excel food sales analyzer
103
 
104
+ Rules:
105
+ 1. Always try to use a tool if relevant.
106
+ 2. Return ONLY the final answer in the requested format.
107
+ 3. Do not guess. If a tool fails, say "Tool not available".
108
+ 4. Follow formats strictly: comma-separated lists, numeric values, chess notation, names only, etc.
109
+ 5. Avoid all explanation unless requested.
110
+ """
111
  )
112
 
113
  def answer_question_sync(question: str) -> str: