include excel reader and audio tools

.gitignore

@@ -26,3 +26,7 @@ config.yaml
 # 6) Any Docker or Kubernetes local files
 docker-compose.override.yml
 *.log
+
+# 7) Test files
+test_sales.xlsx
+test.wav

fetch_gaia_audio.py

@@ -0,0 +1,76 @@
+# fetch_gaia_audio.py
+
+import os
+import re
+import requests
+
+DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
+OUT_PATH = "/mnt/data/test.wav"
+
+def main():
+    # 1) Fetch GAIA questions
+    resp = requests.get(f"{DEFAULT_API_URL}/questions", timeout=15)
+    resp.raise_for_status()
+    questions = resp.json()
+
+    # 2) Try attachments field first
+    for q in questions:
+        for field in ("attachments", "attachment", "audio"):
+            urls = q.get(field)
+            if not urls:
+                continue
+            if isinstance(urls, str):
+                urls = [urls]
+            for url in urls:
+                if is_media_url(url):
+                    return download_audio(url)
+
+    # 3) Fallback: regex scan in question text
+    pattern = re.compile(r"(https?://\S+\.(?:mp3|wav))", re.IGNORECASE)
+    for q in questions:
+        text = q.get("question", "")
+        match = pattern.search(text)
+        if match:
+            url = match.group(1)
+            return download_audio(url)
+
+    print("⚠️  No .mp3/.wav URL found in GAIA payload; skipping download.")
+    return
+
+def is_media_url(url: str) -> bool:
+    return bool(re.match(r"^https?://.*\.(?:mp3|wav)$", url, re.IGNORECASE))
+
+def download_audio(url: str):
+    print(f"Downloading audio from {url}")
+    r = requests.get(url, timeout=30)
+    r.raise_for_status()
+
+    ext = os.path.splitext(url)[1].lower()
+    content = r.content
+
+    if ext == ".mp3":
+        # try to convert to wav if pydub installed
+        try:
+            from pydub import AudioSegment
+            mp3_path = "/mnt/data/tmp.mp3"
+            with open(mp3_path, "wb") as f:
+                f.write(content)
+            audio = AudioSegment.from_mp3(mp3_path)
+            audio.export(OUT_PATH, format="wav")
+            print(f"✔ Saved WAV to {OUT_PATH}")
+            return
+        except ImportError:
+            # fallback: write raw mp3 bytes
+            OUT = OUT_PATH.replace(".wav", ".mp3")
+            with open(OUT, "wb") as f:
+                f.write(content)
+            print(f"⚠ pydub not installed; saved MP3 to {OUT}")
+            return
+
+    # if it's .wav or any other, write directly
+    with open(OUT_PATH, "wb") as f:
+        f.write(content)
+    print(f"✔ Saved WAV to {OUT_PATH}")
+
+if __name__ == "__main__":
+    main()

langgraph3.py

@@ -0,0 +1,143 @@
+import os
+from dotenv import load_dotenv
+import pandas as pd
+import whisper
+
+from langchain_openai import ChatOpenAI
+from langchain_core.messages import SystemMessage, HumanMessage, AIMessage, AnyMessage
+from langchain_core.tools import tool
+from langchain_community.tools.tavily_search import TavilySearchResults
+from langchain_community.document_loaders import WikipediaLoader
+
+from langgraph.graph import StateGraph, MessagesState, START, END
+from langgraph.prebuilt import ToolNode, tools_condition
+
+load_dotenv()
+
+# ─────────────────────────────────────────────
+# System prompt with placeholder for Excel summary
+# ─────────────────────────────────────────────
+SYSTEM_TEMPLATE = """
+You are a razor‑sharp QA agent that answers in **one bare line**.
+- Use tools if factual lookup, audio, or Excel data is needed.
+- Excel data summary is available below.
+- Numbers only for counts.
+- Comma‑separated lists (alphabetize if asked).
+- Codes (IOC, country, etc.) bare.
+- Never apologize or explain.
+Begin.
+
+Excel summary:
+{excel_summary}
+""".strip()
+
+# ─────────────────────────────────────────────
+# TOOLS
+# ─────────────────────────────────────────────
+
+@tool
+def web_search(query: str) -> dict:
+    """Search Tavily for a query and return up to 3 results."""
+    docs = TavilySearchResults(max_results=3).run(query)
+    return {"web_results": "\n".join(d["content"] for d in docs)}
+
+@tool
+def wiki_search(query: str) -> dict:
+    """Search Wikipedia for a query and return up to 2 pages."""
+    pages = WikipediaLoader(query=query, load_max_docs=2).load()
+    return {"wiki_results": "\n\n".join(p.page_content for p in pages)}
+
+@tool
+def transcribe_audio(path: str) -> dict:
+    """Given a local audio file path, return its transcript."""
+    model = whisper.load_model("base")
+    result = model.transcribe(path)
+    return {"transcript": result["text"]}
+
+@tool
+def read_excel(path: str, sheet_name: str = None, sample_rows: int = 5) -> dict:
+    """
+    Read Excel file and return a text summary:
+    - Columns
+    - Sample rows (up to sample_rows)
+    - Basic data types and row count
+    """
+    df = pd.read_excel(path, sheet_name=sheet_name or 0)
+    if isinstance(df, dict):
+        df = next(iter(df.values()))
+    sample = df.head(sample_rows)
+    summary_lines = [
+        f"Columns: {', '.join(df.columns)}",
+        "Data types: " + ", ".join(f"{col}: {dtype}" for col, dtype in df.dtypes.items()),
+        "Sample data:\n" + sample.to_csv(index=False),
+        f"Total rows: {len(df)}"
+    ]
+    return {"excel_summary": "\n".join(summary_lines)}
+
+TOOLS = [web_search, wiki_search, transcribe_audio, read_excel]
+
+# ─────────────────────────────────────────────
+# Load Excel summary ONCE before building system prompt
+# ─────────────────────────────────────────────
+EXCEL_PATH = "test_sales.xlsx"
+excel_summary = read_excel.invoke({"path": EXCEL_PATH})["excel_summary"]
+
+# Build system message with injected Excel summary
+SYSTEM = SystemMessage(content=SYSTEM_TEMPLATE.format(excel_summary=excel_summary))
+
+# ─────────────────────────────────────────────
+# LLM + GRAPH SETUP
+# ─────────────────────────────────────────────
+
+llm = ChatOpenAI(model="gpt-4o-mini", temperature=0.0)
+llm_with_tools = llm.bind_tools(TOOLS)
+
+builder = StateGraph(MessagesState)
+
+def assistant(state: dict) -> dict:
+    msgs = state.get("messages", [])
+    # Ensure system prompt is present at the start
+    if not msgs or not isinstance(msgs[0], SystemMessage):
+        msgs = [SYSTEM] + msgs
+
+    # Let LLM + tools framework handle tool invocation dynamically
+    out: AnyMessage = llm_with_tools.invoke(msgs)
+
+    if isinstance(out, AIMessage) and out.usage_metadata is None:
+        out.usage_metadata = {"input_tokens":0,"output_tokens":0,"total_tokens":0}
+    return {"messages": msgs + [out]}
+
+builder.add_node("assistant", assistant)
+builder.add_node("tools", ToolNode(TOOLS))
+
+builder.add_edge(START, "assistant")
+builder.add_conditional_edges(
+    "assistant",
+    tools_condition,
+    {"tools": "tools", END: END}
+)
+builder.add_edge("tools", "assistant")
+
+graph = builder.compile()
+
+# ─────────────────────────────────────────────
+# Mermaid diagram
+# ───���─────────────────────────────────────────
+print("\n🔍 Mermaid diagram:")
+print(graph.get_graph().draw_mermaid())
+
+# ─────────────────────────────────────────────
+# Smoke test with multi-type questions
+# ─────────────────────────────────────────────
+if __name__ == "__main__":
+    print("🔹 Smoke-testing QA agent")
+    questions = [
+        "How much is 2 + 2?",
+        "What is the capital of France?",
+        "How many rows belong to the food category in the Excel file?",
+        "Which country had the fewest athletes at the 1928 Olympics? Give the IOC code."
+    ]
+    for q in questions:
+        res = graph.invoke({"messages": [HumanMessage(content=q)]})
+        ans = res['messages'][-1].content.strip().rstrip('.')
+        print(f"Q: {q}\n→ A: {ans!r}\n")

requirements.txt

@@ -38,3 +38,5 @@ hf-xet~=1.1.1
 langchain-openai
 tenacity
 openai
+openai-whisper
+openpyxl