added validator node

agent.py

@@ -15,7 +15,6 @@ from langchain_core.messages import SystemMessage
 
 from langgraph.prebuilt import ToolNode, tools_condition
 
-SUPPORTING_FILES_URL = "https://huggingface.co/datasets/gaia-benchmark/GAIA/resolve/main/2023/validation/"
 
 system_prompt = """You are a general AI assistant. I will ask you a question.
 
@@ -60,6 +59,89 @@ def _is_url(path_or_url: str) -> bool:
     except:
         return False
 
+_ARTICLES = {"a", "an", "the"}
+
+def _sanitize_visible_answer(text: str) -> str:
+    """Keep a single-line final answer; strip quotes and leftover tags."""
+    if not text:
+        return ""
+    t = text.strip()
+
+    if (t.startswith('"') and t.endswith('"')) or (t.startswith("'") and t.endswith("'")):
+        t = t[1:-1].strip()
+
+    lines = [ln.strip() for ln in t.splitlines() if ln.strip()]
+    if lines:
+        t = lines[-1]
+
+    t = t.replace("[YOUR FINAL ANSWER]", "").strip()
+    t = t.replace("Final answer: ", "").strip()
+
+
+    t = re.sub(r"\s+", " ", t)
+    t = re.sub(r"<[^>]*>", "", t)
+
+    return t
+
+
+def _is_number_token(s: str) -> bool:
+    return bool(re.fullmatch(r"-?\d+(\.\d+)?", s))
+
+
+def _has_units(s: str) -> bool:
+    return bool(re.search(r"\d\s*[A-Za-z%$]", s))
+
+
+def _has_commas_in_number(s: str) -> bool:
+    return bool(re.search(r"\d,\d", s))
+
+
+def _starts_with_article(s: str) -> bool:
+    toks = re.split(r"[,\s]+", s.strip())
+    return bool(toks) and toks[0].lower() in _ARTICLES
+
+
+def _is_valid_final_answer(ans: str) -> bool:
+    """Validate against your rules:
+       - single line, non-empty
+       - if numeric → no commas, no units
+       - if list → each element validated as number or string
+       - string → no leading article
+    """
+    if not ans or "\n" in ans:
+        return False
+
+    if "," in ans:
+        parts = [p.strip() for p in ans.split(",")]
+        if any(not p for p in parts):
+            return False
+        for p in parts:
+            if re.fullmatch(r".*\d.*", p):  # contains a digit → treat as a number-like
+                if not _is_number_token(p):
+                    return False
+                if _has_commas_in_number(p):
+                    return False
+                if _has_units(p):
+                    return False
+            else:
+                if _starts_with_article(p):
+                    return False
+        return True
+
+    if re.fullmatch(r".*\d.*", ans):  # number-like
+        if not _is_number_token(ans):
+            return False
+        if _has_commas_in_number(ans):
+            return False
+        if _has_units(ans):
+            return False
+        return True
+    else:
+        if _starts_with_article(ans):
+            return False
+        return True
+
+
 def _process_uploaded_file(file_name: str, file_path: str) -> str:
     """Process a single local file or file URL and return context for the question."""
     try:
@@ -101,7 +183,6 @@ def build_and_compile():
     python_code,
     image_info,
     read_mp3_transcript,
-    pdf_text_extractor,
     ocr_image,
     math_solver,
     plot_data_tool,
@@ -121,6 +202,7 @@ def build_and_compile():
 
     llm = init_chat_model("openai:gpt-4.1-mini",temperature=0, seed=42)
     llm_with_tools = llm.bind_tools(tools)
+    final_llm = llm.bind(response_format={"type": "json_object"})
 
     def chatbot(state: State):
         file_context = ""
@@ -129,18 +211,72 @@ def build_and_compile():
         final_prompt = system_prompt + file_context
         return {"messages": [llm_with_tools.invoke([SystemMessage(final_prompt)] + state["messages"])]}
 
+    def validator(state: State):
+        """
+        Ensure the last assistant message is a valid final answer per system rules.
+        If invalid, rewrite once with final_llm (JSON) and output only final_answer.
+        """
+        # Get last assistant message text
+        last = state["messages"][-1]
+        text = getattr(last, "content", "") or str(last)
 
-    graph_builder.add_node("chatbot", chatbot)
+        # 1) sanitize
+        clean = _sanitize_visible_answer(text)
+
+        # 2) validate
+        if _is_valid_final_answer(clean):
+            # Replace the last message with the sanitized one-line answer
+            return {"messages": [{"role": "assistant", "content": clean}]}
 
+        # 3) one-shot fixer pass (no tools, JSON enforced)
+        fix_instruction = (
+            "Rewrite the final answer to comply with these rules:\n"
+            "- Output only the final answer (single line), no extra words.\n"
+            "- Numbers should always be expressed as digits.\n"
+            "- If number: no commas, no units.\n"
+            "- If string: no leading articles ('a','an','the'); no abbreviations.\n"
+            "- If list: comma-separated; apply the same rules to each element.\n\n"
+            "Return JSON: {\"final_answer\": \"...\"}."
+        )
+        msgs = [
+            SystemMessage(system_prompt),
+            {"role": "user", "content": fix_instruction + f"\n\nOriginal answer:\n{clean}"}
+        ]
+        fixed = final_llm.invoke(msgs)
+        fixed_text = str(getattr(fixed, "content", "") or "").strip()
+        try:
+            obj = json.loads(fixed_text)
+            fa = (obj.get("final_answer") or "").strip()
+        except Exception:
+            # fallback: keep sanitized original if JSON parsing fails
+            fa = clean
+
+        fa = _sanitize_visible_answer(fa)
+        if not _is_valid_final_answer(fa):
+            # last resort: keep last line of whatever we have
+            fa = (fa or clean).splitlines()[-1].strip()
+
+        return {"messages": [{"role": "assistant", "content": fa}]}
+
+    graph_builder.add_node("chatbot", chatbot)
     tool_node = ToolNode(tools=tools)
     graph_builder.add_node("tools", tool_node)
+    graph_builder.add_node("validator", validator)
 
+    # If the model wants to call tools → go to tools; else → go to validator
     graph_builder.add_conditional_edges(
         "chatbot",
         tools_condition,
+        {"tools": "tools", "__end__": "validator"},
     )
-    # Any time a tool is called, we return to the chatbot to decide the next step
+
+    # After tools run, go back to chatbot
     graph_builder.add_edge("tools", "chatbot")
+
+    # After validator, we are done
+    graph_builder.add_edge("validator", END)
+
     graph_builder.add_edge(START, "chatbot")
+
     graph = graph_builder.compile()
     return graph

tools.py

@@ -211,37 +211,6 @@ def read_mp3_transcript(path: str) -> str:
         return _fmt_error("read_mp3_transcript", e)
 
 
-@tool("pdf_text_extractor")
-def pdf_text_extractor(args: str) -> str:
-    """Extract text from a PDF. Usage:
-    - 'path/to/file.pdf'
-    - 'path/to/file.pdf|pages=1-3' (1-indexed inclusive range)
-    Returns a concatenated text excerpt (truncated)."""
-    try:
-        if pdfplumber is None:
-            raise RuntimeError("pdfplumber not installed")
-        path, start, end = args, None, None
-        m = re.search(r"\|pages=(\d+)-(\d+)$", args.strip())
-        if m:
-            path = args[: args.rfind("|pages=")]
-            start, end = int(m.group(1)), int(m.group(2))
-        text_parts: List[str] = []
-        with pdfplumber.open(path) as pdf:
-            total = len(pdf.pages)
-            s = max(1, start) if start else 1
-            e = min(end, total) if end else total
-            for p in range(s - 1, e):
-                page = pdf.pages[p]
-                text_parts.append(page.extract_text() or "")
-        text = "\n".join(text_parts).strip()
-        if not text:
-            text = "(no extractable text)"
-        meta = {"path": path, "pages": f"{start or 1}-{end or 'end'}"}
-        return _fmt_block("PDFText", meta, _truncate(text, 4000))
-    except Exception as e:
-        return _fmt_error("pdf_text_extractor", e)
-
-
 @tool("ocr_image")
 def ocr_image(path: str) -> str:
     """Run OCR on an image and return extracted text (requires pytesseract + Tesseract installed)."""