Using LangChain with DuckDuckGO Wikipedia & PythonREPL Tools

LangChain agents can be made to choose among three built-in tools—Wikipedia, DuckDuckGo search, and Python’s Read-Evaluate-Print Loop (Python REPL)—based on the question being asked. The practical payoff is an agent that can look up factual information on the web, pull structured background from Wikipedia, and run real computations in Python, switching tools automatically instead of forcing a single workflow.

The setup starts with installing the Wikipedia API wrapper and then wiring each tool into LangChain as a “chain” or callable function. The Wikipedia tool is configured so a query like “LangChain” returns the matching Wikipedia page content, while the Python REPL tool evaluates Python code and prints results (for example, computing 17 * 2 to produce 34). DuckDuckGo is added as a free search option—unlike paid SERP APIs—so the agent can retrieve current or general web information when Wikipedia isn’t the best fit.

Those tools are then bundled into a list and handed to a zero-shot agent using the ZeroShotAgent with the ZeroShot React description prompt template. The agent runs with verbose logging and a max-iteration limit, meaning it can “think” through multiple steps—deciding whether to call a tool, inspect the observation returned, and then either call another tool or produce a final answer.

Concrete examples show the tool-selection behavior. For “When was Barack Obama born?”, the agent routes the task to DuckDuckGo search, extracts the birth date from the returned results, and then answers directly. For “17 by 6,” it recognizes a computation task and uses Python REPL to calculate the result (102) rather than searching the web.

The prompt itself is central: it instructs the agent to answer questions “as best you can” using the available tools, and it provides tool descriptions that act like decision rules. Wikipedia is described as useful for looking up a “topic, country or person,” DuckDuckGo is described as useful for internet search when another tool can’t find the needed information, and Python REPL is described as the way to run Python code for calculations.

When asked “tell me about langChain,” the agent chooses DuckDuckGo instead of Wikipedia because the two lookup tools overlap. However, the Wikipedia path becomes more likely when the question matches the Wikipedia description more clearly. For “tell me about Singapore,” the agent selects Wikipedia, retrieves the page, and then narrows the response to Singapore as a country (an island and city-state in Southeast Asia).

The agent also demonstrates mixed tool use under uncertainty. For “is 11 a prime number?”, Python REPL initially produces an inconclusive “might be prime” style result, prompting the agent to fall back to DuckDuckGo to confirm. Rephrasing the task by asking it to write a Python function to test primality yields a cleaner multi-step solution: the agent generates the function, runs it, and returns that 11 is prime.

Overall, the workflow shows how tool descriptions and prompt structure drive an agent’s routing decisions, and how better prompting (or adding examples) can improve reliability—especially when tool outputs are ambiguous or when multiple tools can answer the same kind of question.