GPT-4 Turbo with Google Web Browsing (Assistants API)

A practical Assistants API pattern is on display: rewrite a user question into a Google-friendly search query, fetch fresh web results, scrape the pages for relevant text, then feed that “grounded” context back into GPT-4 Turbo so answers reflect current information instead of relying on training-time knowledge. The workflow matters because it turns a general-purpose model into a news-and-facts retriever that can cite up-to-date details—illustrated with very recent claims about Sam Altman and with a sports result example.

The system starts when a user types a question. A first GPT-4 Turbo call reformulates the query into a shorter, search-optimized version (for example, converting “Sam Altman fired from OpenAI” into a Google-ready query). That rewritten query is sent to an Assistants API flow that uses function calling: one tool queries Google via SerpAPI to obtain organic result URLs, and another tool scrapes the returned pages using BeautifulSoup 4 to extract the relevant text. The pipeline then combines the original user question with the scraped context and makes a second GPT-4 Turbo call to generate the final answer.

A key design choice is how the final response is constrained. A system message instructs the model to return only the essential parts that answer the user’s original question, while adding three bullet points that justify the response using the grounded text. In testing, the approach produces answers that track the retrieved material—for instance, for the query about whether Sam Altman was fired, the model returns a summary consistent with the scraped sources, including details like timing (“fired… Friday”) and downstream effects (uncertainty and resignations). The presenter emphasizes that this kind of grounding is not dependent on the model’s training data, since the events referenced were described as occurring only days earlier.

The same mechanism is demonstrated with a non-governance question: “who won the Las Vegas F1 Grand Prix.” After query rewriting and scraping, GPT-4 Turbo generates a winner and supporting details (including mention of a penalty and a collision) drawn from the retrieved page text. The system also supports output-format experimentation. By changing the instruction prompt, the assistant can return structured JSON (e.g., a concise answer field) or even a short poem, while still using the scraped context as the factual basis.

Implementation details reinforce the pattern. The code defines separate functions for (1) generating the Google search query with GPT-4 Turbo, (2) retrieving organic results via SerpAPI, and (3) scraping each URL with BeautifulSoup. Function calling is used serially rather than in parallel because scraping requires the URLs first. The workflow is then run in a terminal environment; the OpenAI Playground is noted as unable to scrape websites in this setup. Overall, the transcript presents a working blueprint for “web-grounded” Q&A using GPT-4 Turbo plus Assistants API tool calls, with flexible output formatting and a clear grounding strategy.