Replace SemanticScholarTool with WikipediaTool in notebook

jeremymanning · claude · jeremymanning · commit 9315e8c1c227 · 2026-03-03T15:26:21.000-05:00
Semantic Scholar API was rate-limiting (429 errors) during demos.
Wikipedia API is free, reliable, and needs no authentication —
much better for a classroom notebook.

Updated all references: Part 2 custom tool, Part 3 comparison,
exercise suggestions, and summary table.

Co-Authored-By: Claude Opus 4.6 &lt;noreply@anthropic.com&gt;
diff --git a/slides/week9/agents_demo.ipynb b/slides/week9/agents_demo.ipynb
@@ -127,116 +127,26 @@
   {
    "cell_type": "markdown",
    "metadata": {},
-   "source": [
-    "---\n",
-    "\n",
-    "## Part 2: Custom tools\n",
-    "\n",
-    "The real power of agents comes from giving them **custom tools**. A tool is just a Python function with a description that tells the LLM when and how to use it.\n",
-    "\n",
-    "Let's build a tool that looks up papers on the [Semantic Scholar](https://www.semanticscholar.org/) API — a real academic search engine."
-   ]
+   "source": "---\n\n## Part 2: Custom tools\n\nThe real power of agents comes from giving them **custom tools**. A tool is just a Python function with a description that tells the LLM when and how to use it.\n\nLet's build a tool that looks up articles on [Wikipedia](https://en.wikipedia.org/) — using their free, public API."
   },
   {
    "cell_type": "code",
    "execution_count": null,
    "metadata": {},
    "outputs": [],
-   "source": [
-    "from smolagents import Tool\n",
-    "import requests\n",
-    "\n",
-    "\n",
-    "class SemanticScholarTool(Tool):\n",
-    "    \"\"\"Search for academic papers using the Semantic Scholar API.\"\"\"\n",
-    "    name = \"paper_search\"\n",
-    "    description = (\n",
-    "        \"Search for academic research papers by topic. Returns titles, authors, \"\n",
-    "        \"year, citation count, and a URL for each paper. Use this when you need \"\n",
-    "        \"to find scientific literature or verify academic claims.\"\n",
-    "    )\n",
-    "    inputs = {\n",
-    "        \"query\": {\n",
-    "            \"type\": \"string\",\n",
-    "            \"description\": \"The search query (e.g., 'chain of thought prompting')\"\n",
-    "        },\n",
-    "        \"max_results\": {\n",
-    "            \"type\": \"integer\",\n",
-    "            \"description\": \"Maximum number of papers to return (default: 5)\",\n",
-    "            \"nullable\": True\n",
-    "        }\n",
-    "    }\n",
-    "    output_type = \"string\"\n",
-    "\n",
-    "    def forward(self, query: str, max_results: int = 5) -> str:\n",
-    "        url = \"https://api.semanticscholar.org/graph/v1/paper/search\"\n",
-    "        params = {\n",
-    "            \"query\": query,\n",
-    "            \"limit\": min(max_results or 5, 10),\n",
-    "            \"fields\": \"title,authors,year,citationCount,url\"\n",
-    "        }\n",
-    "        try:\n",
-    "            resp = requests.get(url, params=params, timeout=10)\n",
-    "            resp.raise_for_status()\n",
-    "            data = resp.json()\n",
-    "        except Exception as e:\n",
-    "            return f\"Error searching papers: {e}\"\n",
-    "\n",
-    "        papers = data.get(\"data\", [])\n",
-    "        if not papers:\n",
-    "            return \"No papers found for this query.\"\n",
-    "\n",
-    "        results = []\n",
-    "        for p in papers:\n",
-    "            authors = \", \".join(a[\"name\"] for a in (p.get(\"authors\") or [])[:3])\n",
-    "            if len(p.get(\"authors\", [])) > 3:\n",
-    "                authors += \" et al.\"\n",
-    "            results.append(\n",
-    "                f\"- **{p['title']}** ({p.get('year', '?')})\\n\"\n",
-    "                f\"  Authors: {authors}\\n\"\n",
-    "                f\"  Citations: {p.get('citationCount', '?')} | \"\n",
-    "                f\"URL: {p.get('url', 'N/A')}\"\n",
-    "            )\n",
-    "        return \"\\n\\n\".join(results)\n",
-    "\n",
-    "\n",
-    "# Test the tool directly (before giving it to an agent)\n",
-    "paper_tool = SemanticScholarTool()\n",
-    "print(paper_tool.forward(\"ReAct reasoning acting language models\", max_results=3))"
-   ]
+   "source": "from smolagents import Tool\nimport requests\n\n\nclass WikipediaTool(Tool):\n    \"\"\"Look up a topic on Wikipedia and return a summary.\"\"\"\n    name = \"wiki_lookup\"\n    description = (\n        \"Look up a topic on Wikipedia and return a concise summary. Use this \"\n        \"when you need factual background information about a person, place, \"\n        \"concept, or event. Returns the article title, a summary, and a URL.\"\n    )\n    inputs = {\n        \"query\": {\n            \"type\": \"string\",\n            \"description\": \"The topic to search for (e.g., 'chain of thought prompting')\"\n        }\n    }\n    output_type = \"string\"\n\n    def forward(self, query: str) -> str:\n        # Step 1: Search for matching articles\n        search_url = \"https://en.wikipedia.org/w/api.php\"\n        search_params = {\n            \"action\": \"query\",\n            \"list\": \"search\",\n            \"srsearch\": query,\n            \"srlimit\": 3,\n            \"format\": \"json\"\n        }\n        try:\n            resp = requests.get(search_url, params=search_params, timeout=10)\n            resp.raise_for_status()\n            results = resp.json()[\"query\"][\"search\"]\n        except Exception as e:\n            return f\"Error searching Wikipedia: {e}\"\n\n        if not results:\n            return f\"No Wikipedia articles found for '{query}'.\"\n\n        # Step 2: Get the summary of the top result\n        title = results[0][\"title\"]\n        summary_url = f\"https://en.wikipedia.org/api/rest_v1/page/summary/{requests.utils.quote(title)}\"\n        try:\n            resp = requests.get(summary_url, timeout=10)\n            resp.raise_for_status()\n            data = resp.json()\n        except Exception as e:\n            return f\"Error fetching summary for '{title}': {e}\"\n\n        extract = data.get(\"extract\", \"No summary available.\")\n        page_url = data.get(\"content_urls\", {}).get(\"desktop\", {}).get(\"page\", \"N/A\")\n\n        return (\n            f\"**{data.get('title', title)}**\\n\\n\"\n            f\"{extract}\\n\\n\"\n            f\"Read more: {page_url}\"\n        )\n\n\n# Test the tool directly (before giving it to an agent)\nwiki_tool = WikipediaTool()\nprint(wiki_tool.forward(\"ReAct framework language models\"))"
   },
   {
    "cell_type": "code",
    "execution_count": null,
    "metadata": {},
    "outputs": [],
-   "source": [
-    "# Now give the agent BOTH tools: web search AND paper search\n",
-    "research_agent = CodeAgent(\n",
-    "    tools=[search_tool, paper_tool],\n",
-    "    model=model,\n",
-    "    max_steps=6,\n",
-    "    verbosity_level=2\n",
-    ")\n",
-    "\n",
-    "result = research_agent.run(\n",
-    "    \"Find the most cited paper on chain-of-thought prompting. \"\n",
-    "    \"Who are the authors and how many citations does it have?\"\n",
-    ")\n",
-    "print(f\"\\nFinal answer: {result}\")"
-   ]
+   "source": "# Now give the agent BOTH tools: web search AND Wikipedia lookup\nresearch_agent = CodeAgent(\n    tools=[search_tool, wiki_tool],\n    model=model,\n    max_steps=6,\n    verbosity_level=2\n)\n\nresult = research_agent.run(\n    \"What is chain-of-thought prompting? Look it up on Wikipedia, \"\n    \"then search the web for who invented it.\"\n)\nprint(f\"\\nFinal answer: {result}\")"
   },
   {
    "cell_type": "markdown",
    "metadata": {},
-   "source": [
-    "### 💡 Discussion\n",
-    "\n",
-    "- The agent had two tools available. How did it decide which one to use?\n",
-    "- Look at the tool `description` field. How does this influence the agent's behavior?\n",
-    "- What would happen if two tools had very similar descriptions?\n",
-    "- Try changing the description to something vague. Does the agent still pick the right tool?"
-   ]
+   "source": "### 💡 Discussion\n\n- The agent had two tools available. How did it decide which one to use?\n- Look at the tool `description` field. How does this influence the agent's behavior?\n- What would happen if two tools had very similar descriptions?\n- Try changing the description to something vague — like `\"Does stuff\"`. Does the agent still pick the right tool?"
   },
   {
    "cell_type": "markdown",
@@ -263,36 +173,7 @@
    "execution_count": null,
    "metadata": {},
    "outputs": [],
-   "source": [
-    "from smolagents import ToolCallingAgent\n",
-    "\n",
-    "# Same model, same tools — different agent type\n",
-    "tool_calling_agent = ToolCallingAgent(\n",
-    "    tools=[search_tool, paper_tool],\n",
-    "    model=model,\n",
-    "    max_steps=6,\n",
-    "    verbosity_level=2\n",
-    ")\n",
-    "\n",
-    "# Ask both agents the same question\n",
-    "question = \"What is MCP (Model Context Protocol) and who created it?\"\n",
-    "\n",
-    "print(\"=\" * 60)\n",
-    "print(\"CODE AGENT\")\n",
-    "print(\"=\" * 60)\n",
-    "code_result = agent.run(question)\n",
-    "\n",
-    "print(\"\\n\" + \"=\" * 60)\n",
-    "print(\"TOOL-CALLING AGENT\")\n",
-    "print(\"=\" * 60)\n",
-    "tc_result = tool_calling_agent.run(question)\n",
-    "\n",
-    "print(\"\\n\" + \"=\" * 60)\n",
-    "print(\"COMPARISON\")\n",
-    "print(\"=\" * 60)\n",
-    "print(f\"Code agent answer: {code_result}\")\n",
-    "print(f\"Tool-calling agent answer: {tc_result}\")"
-   ]
+   "source": "from smolagents import ToolCallingAgent\n\n# Same model, same tools — different agent type\ntool_calling_agent = ToolCallingAgent(\n    tools=[search_tool, wiki_tool],\n    model=model,\n    max_steps=6,\n    verbosity_level=2\n)\n\n# Ask both agents the same question\nquestion = \"What is MCP (Model Context Protocol) and who created it?\"\n\nprint(\"=\" * 60)\nprint(\"CODE AGENT\")\nprint(\"=\" * 60)\ncode_result = agent.run(question)\n\nprint(\"\\n\" + \"=\" * 60)\nprint(\"TOOL-CALLING AGENT\")\nprint(\"=\" * 60)\ntc_result = tool_calling_agent.run(question)\n\nprint(\"\\n\" + \"=\" * 60)\nprint(\"COMPARISON\")\nprint(\"=\" * 60)\nprint(f\"Code agent answer: {code_result}\")\nprint(f\"Tool-calling agent answer: {tc_result}\")"
   },
   {
    "cell_type": "markdown",
@@ -608,20 +489,7 @@
   {
    "cell_type": "markdown",
    "metadata": {},
-   "source": [
-    "---\n",
-    "\n",
-    "## Part 6: Build your own agent (exercise)\n",
-    "\n",
-    "Now it's your turn! Create a custom tool and use it in an agent. Here are some ideas:\n",
-    "\n",
-    "1. **Wikipedia lookup tool** — Use the Wikipedia API to fetch article summaries\n",
-    "2. **Unit converter** — Convert between metric and imperial units\n",
-    "3. **Sentiment analyzer** — Analyze the sentiment of text passages\n",
-    "4. **Course schedule tool** — Look up when classes meet\n",
-    "\n",
-    "Use the `SemanticScholarTool` above as a template."
-   ]
+   "source": "---\n\n## Part 6: Build your own agent (exercise)\n\nNow it's your turn! Create a custom tool and use it in an agent. Here are some ideas:\n\n1. **Weather tool** — Use the [Open-Meteo API](https://open-meteo.com/) (free, no key) to get current weather for a city\n2. **Unit converter** — Convert between metric and imperial units\n3. **Sentiment analyzer** — Analyze the sentiment of text passages\n4. **Course schedule tool** — Look up when classes meet\n\nUse the `WikipediaTool` above as a template."
   },
   {
    "cell_type": "code",
@@ -654,33 +522,7 @@
   {
    "cell_type": "markdown",
    "metadata": {},
-   "source": [
-    "---\n",
-    "\n",
-    "## Summary\n",
-    "\n",
-    "| Concept | What we built | Key takeaway |\n",
-    "|---------|--------------|-------------|\n",
-    "| **ReAct loop** | Agent with web search | LLMs can reason step-by-step, calling tools as needed |\n",
-    "| **Custom tools** | Semantic Scholar search | Tools are just functions with good descriptions |\n",
-    "| **Code vs. tool-calling** | Compared both paradigms | Code agents are flexible but riskier; tool-calling is safer |\n",
-    "| **RAG agent** | Knowledge base search | Agents can ground answers in authoritative sources |\n",
-    "| **Safety** | Injection demo, step limits | Autonomy must be matched to stakes (lecture principle) |\n",
-    "\n",
-    "## Connections to the lecture\n",
-    "\n",
-    "- **Slide 5 (ReAct)**: You saw the Thought → Action → Observation loop in action\n",
-    "- **Slide 7 (MCP)**: Our tools follow the same pattern as MCP — structured descriptions that any model can use\n",
-    "- **Slide 8 (SWE-bench)**: Real coding agents use the same loop, just with more powerful tools\n",
-    "- **Slides 14–16 (Safety)**: We demonstrated prompt injection, step limits, and the trust boundary\n",
-    "\n",
-    "## Further exploration\n",
-    "\n",
-    "- [**smolagents documentation**](https://huggingface.co/docs/smolagents/index) — Full guide to building agents\n",
-    "- [**HuggingFace Agents Course**](https://huggingface.co/learn/agents-course) — Free course on building AI agents\n",
-    "- [**LangChain**](https://www.langchain.com/) — The most popular agent framework (more complex, more features)\n",
-    "- [**Yao et al. (2023)**](https://arxiv.org/abs/2210.03629) — The original ReAct paper"
-   ]
+   "source": "---\n\n## Summary\n\n| Concept | What we built | Key takeaway |\n|---------|--------------|-------------|\n| **ReAct loop** | Agent with web search | LLMs can reason step-by-step, calling tools as needed |\n| **Custom tools** | Wikipedia lookup | Tools are just functions with good descriptions |\n| **Code vs. tool-calling** | Compared both paradigms | Code agents are flexible but riskier; tool-calling is safer |\n| **RAG agent** | Knowledge base search | Agents can ground answers in authoritative sources |\n| **Safety** | Injection demo, step limits | Autonomy must be matched to stakes (lecture principle) |\n\n## Connections to the lecture\n\n- **Slide 5 (ReAct)**: You saw the Thought → Action → Observation loop in action\n- **Slide 7 (MCP)**: Our tools follow the same pattern as MCP — structured descriptions that any model can use\n- **Slide 8 (SWE-bench)**: Real coding agents use the same loop, just with more powerful tools\n- **Slides 14–16 (Safety)**: We demonstrated prompt injection, step limits, and the trust boundary\n\n## Further exploration\n\n- [**smolagents documentation**](https://huggingface.co/docs/smolagents/index) — Full guide to building agents\n- [**HuggingFace Agents Course**](https://huggingface.co/learn/agents-course) — Free course on building AI agents\n- [**LangChain**](https://www.langchain.com/) — The most popular agent framework (more complex, more features)\n- [**Yao et al. (2023)**](https://arxiv.org/abs/2210.03629) — The original ReAct paper"
   }
  ],
  "metadata": {
