Update app.py

app.py

@@ -87,74 +87,34 @@ class BasicAgent:
             #description="Analyses images to locate and describe objects in them.",
             #trust_remote_code=True,
         #)
-        """You are a research agent solving GAIA benchmark questions. You must be
-         precise, because answers are graded by EXACT STRING MATCH against a ground
-         truth — no extra words, no units unless explicitly asked, no restating the
-         question, no "FINAL ANSWER:" prefix, no explanation in the final output.
-
-         Work in a strict Thought / Action / Observation cycle for every step:
-         - Thought: reason briefly about what you know, what's missing, and which
-           single tool call gets you closer to the answer.
-         - Action: call exactly one tool with the specific input it needs.
-         - Observation: read the tool's output carefully before your next Thought.
-         Never guess an answer without an Observation to support it. If a tool
-         returns nothing useful, change your query or try a different tool rather
-         than repeating the same call.
-
-         TOOL SELECTION RULES — check in this order for every question:
-
-         1. FILES: If the question includes a task_id and references a file, image,
-           spreadsheet, or document, your first Action must be fetch_task_file(task_id)
-           to download it before anything else.
-           - If the downloaded file ends in .csv or .xlsx: use read_spreadsheet to
-           inspect it as a table before reasoning about numbers or entries.
-           - If it ends in .pdf: use pdf_tool to extract its text.
-           - If it's an image: describe what you need to find, and reason only from
-           details you can verify — do not invent visual details you cannot confirm.
-
-           2. YOUTUBE: If the question contains a youtube.com or youtu.be URL, use
-              get_youtube_transcript on that URL. Read the full transcript before
-              answering — do not assume content from the title alone.
-
-          3. OTHER URLS: If the question contains any other URL, use visit_webpage on
-             it directly. Do not web_search for a page you were already given a link to.
-
-          4. GENERAL KNOWLEDGE OR CURRENT INFO: If there is no file and no URL, or the
-             given source doesn't contain the answer, use web_search with a short,
-             specific query (not the full question text verbatim). If the first
-             result set doesn't answer it, refine the query — try alternate terms,
-             dates, or named entities — before giving up.
-
-          5. MATH / COUNTING / DATA MANIPULATION: Do these directly in Python code
-             within your Action step rather than trying to compute them mentally or
-             estimate them.
-
-          ANSWER FORMAT:
-            - Once you have verified evidence (not assumption) for the answer, stop.
-            - Return only the answer itself: a number, a name, a short string, or a
-              list, in the exact form the question asks for (e.g. "3" not "three
-              albums", "Paris" not "The city is Paris").
-            - If after reasonable effort across tools you cannot find a verified
-              answer, return your best-supported guess rather than an empty or vague
-             response — GAIA has no partial credit for admitting uncertainty.
-             """
-        pdf_tool = load_tool(
+        
+        """pdf_tool = load_tool(
             "matterattetatte/pdf-upload-extractor-tool", 
             trust_remote_code = True
-        )
+        )"""
         self.agent = CodeAgent(
             tools=[
                 DuckDuckGoSearchTool(),
                 VisitWebpageTool(),
-                pdf_tool,
+                #pdf_tool,
                 fetch_task_file,
                 read_spreadsheet,
                 get_youtube_transcript,
                 #os.environ["HF_TOKEN"] = "hf_token",
                 #image_analysis_tool,
                 ],
+                additional_authorized_imports=[
+                                  "pypdf",        # basic text extraction
+                                "pdfplumber",   # better for tables/layout inside PDFs
+                                    "fitz",         # PyMuPDF — fast, handles images + text well
+                                    "io",           # often needed for BytesIO when handling file bytes
+                                 "requests",     # for fetch_task_file
+                                  "pandas",       # for read_spreadsheet
+                                 "openpyxl",     # backend pandas needs for .xlsx
+                                   "bs4",          # for parsing HTML from VisitWebpageTool output if needed
+                             ]
             model= InferenceClientModel("Qwen/Qwen2.5-Coder-32B-Instruct"),
-            instructions = "If the question includes a task_id and mentions a file, call fetch_task_file first; route YouTube URLs to get_youtube_transcript, other URLs to visit_webpage, PDFs to pdf_tool, and spreadsheets to read_spreadsheet, using web_search only when no URL or file is given,then respond with only the exact final answer value, no explanation, no prefix.",
+            #instructions = "If the question includes a task_id and mentions a file, call fetch_task_file first; route YouTube URLs to get_youtube_transcript, other URLs to visit_webpage, PDFs to pdf_tool, and spreadsheets to read_spreadsheet, using web_search only when no URL or file is given,then respond with only the exact final answer value, no explanation, no prefix.",
             max_steps=20,
         )
 #answering questions