Update app.py

app.py

@@ -20,8 +20,6 @@ from langchain_core.messages import HumanMessage, SystemMessage
 from langchain_core.tools import tool
 from langchain_groq import ChatGroq
 from langchain_community.tools import DuckDuckGoSearchRun
-from langchain_community.tools.wikipedia.tool import WikipediaQueryRun
-from langchain_community.utilities.wikipedia import WikipediaAPIWrapper
 from langgraph.graph import END, StateGraph
 
 
@@ -119,7 +117,6 @@ def _is_image(ct: str, data: bytes) -> bool:
         return True
     if data.startswith(b"GIF87a") or data.startswith(b"GIF89a"):
         return True
-    # RIFF can also be WAV, so verify WEBP signature.
     if data[:4] == b"RIFF" and data[8:12] == b"WEBP":
         return True
     return False
@@ -144,6 +141,11 @@ def _is_audio_or_video(ct: str, path: str) -> bool:
     return ct.startswith("audio/") or ct.startswith("video/") or ext in AUDIO_VIDEO_EXTS
 
 
+def is_youtube_question(question: str) -> bool:
+    q = question.lower()
+    return "youtube.com/watch" in q or "youtu.be/" in q
+
+
 def detect_local_file_kind(task_id: str) -> tuple[str, str | None]:
     local_path = get_task_file(task_id)
     if not local_path:
@@ -347,11 +349,6 @@ def run_python_file(path: Path, timeout_seconds: int = 8) -> str:
         return f"ERROR: Code execution failed: {type(e).__name__}: {e}"
 
 web_search_tool = DuckDuckGoSearchRun(name="web_search")
-wikipedia_tool = WikipediaQueryRun(
-    name="wikipedia",
-    api_wrapper=WikipediaAPIWrapper(top_k_results=3, doc_content_chars_max=3000),
-)
-
 
 def safe_tool_run(tool_obj: Any, query: str, limit: int = 6000) -> str:
     try:
@@ -398,6 +395,15 @@ def is_bad_answer(answer: str) -> bool:
         "not enough information",
         "unable to answer",
         "no answer",
+        "no answer found",
+        "i could not find",
+        "could not find",
+        "not found",
+        "not in the search results",
+        "this answer is not",
+        "unknown",
+        "insufficient information",
+        "cannot determine",
     ]
     return any(m in a for m in bad_markers)
 
@@ -408,7 +414,7 @@ def reversed_english_question(question: str) -> bool:
     return sum(1 for m in markers if m in rev) >= 2
 
 
-def maybe_solve_directly_with_python(question: str) -> str | None:
+def solve_directly_with_python(question: str) -> str | None:
     q = question.strip()
 
     if reversed_english_question(q):
@@ -465,6 +471,11 @@ def last_nonempty_line(text: str) -> str:
     return lines[-1] if lines else ""
 
 
+def extract_youtube_id(question: str) -> str | None:
+    m = re.search(r"(?:v=|youtu\.be/)([A-Za-z0-9_-]{11})", question)
+    return m.group(1) if m else None
+
+
 class AgentState(TypedDict):
     question: str
     task_id: str
@@ -558,6 +569,8 @@ class BasicAgent:
             route = "solve_direct"
         elif direct_question(question):
             route = "solve_direct"
+        elif is_youtube_question(question):
+            route = "solve_youtube"
         else:
             route = "solve_research"
 
@@ -655,7 +668,7 @@ class BasicAgent:
         file_kind = state.get("file_kind", "none")
         local_path = state.get("local_path")
 
-        shortcut = maybe_solve_directly_with_python(question)
+        shortcut = solve_directly_with_python(question)
         if shortcut is not None:
             return {"context": "Solved by deterministic Python shortcut.", "raw_answer": shortcut}
 
@@ -678,29 +691,60 @@ class BasicAgent:
         print(f"[research query] {query}")
 
         web_results = safe_tool_run(web_search_tool, query, limit=5000)
-        wiki_results = safe_tool_run(wikipedia_tool, query, limit=5000)
 
         print(f"[web len] {len(web_results)}")
-        print(f"[wiki len] {len(wiki_results)}")
         print(f"[web preview] {repr(web_results[:500])}")
-        print(f"[wiki preview] {repr(wiki_results[:500])}")
 
         context = truncate_text(
-            f"Search query:\n{query}\n\n"
-            f"Web search results:\n{web_results}\n\n"
-            f"Wikipedia results:\n{wiki_results}",
+            f"Search query:\n{query}\n\n",
+            f"Web search results:\n{web_results}\n\n",
             MAX_SEARCH_CONTEXT_CHARS,
         )
 
         raw_answer = self.answer_from_context(
             question=question,
             context=context,
-            context_label="Web/Wikipedia research results",
+            context_label="Web research results",
             llm=self.research_llm,
         )
 
         print(f"[research raw_answer] {repr(raw_answer[:500])}")
         return {"context": context, "raw_answer": raw_answer}
+    
+    def solve_youtube(self, state: AgentState) -> dict:
+        question = state["question"]
+        video_id = extract_youtube_id(question)
+
+        queries = []
+
+        if video_id:
+            queries += [
+                f'"{video_id}" transcript',
+                f'"{video_id}" subtitles',
+                f'"{video_id}" "{question[:80]}"',
+            ]
+
+        queries.append(question)
+
+        parts = []
+
+        for query in queries:
+            result = safe_tool_run(web_search_tool, query, limit=5000)
+            parts.append(f"Query: {query}\nResults:\n{result}")
+
+        context = "\n\n---\n\n".join(parts)[:16000]
+
+        raw_answer = self.answer_from_context(
+            question=question,
+            context=context,
+            context_label="YouTube/web transcript search results",
+            llm=self.research_llm,
+        )
+
+        return {
+            "context": context,
+            "raw_answer": raw_answer,
+        }
 
     def verify_answer(self, state: AgentState) -> dict[str, Any]:
         question = state.get("question", "")
@@ -909,7 +953,7 @@ with gr.Blocks() as demo:
         """
         **Architecture:** `classify_task → route_by_type → solve_* → verify_answer → final_cleaner`.
 
-        Local files are routed deterministically by Python. Web/Wikipedia are called only inside `solve_research`, without automatic LLM tool-calling.
+        Local files are routed deterministically by Python. Web are called only inside `solve_research`, without automatic LLM tool-calling.
         """
     )