DataTalks.Club

Articles Slack Events Podcast Wiki Books Courses

Course Wiki

Function Calling — LLM Zoomcamp Module 1

--- video_url: 'https://www.youtube.com/watch?v=CeEki_0mdGo&list=PL3MmuxUbc_hLZFNgSad56pDBKK8KO0XIv' --- # Function Calling

Related Pages

llmz-module-01

LLM Zoomcamp › Module 1: Agentic RAG › Function Calling

Notes

video_url: “https://www.youtube.com/watch?v=CeEki_0mdGo&list=PL3MmuxUbc_hLZFNgSad56pDBKK8KO0XIv”

Function Calling

In the previous lesson we built a RAG pipeline with RAGBase.rag() and saw it fail on the “Olama” typo. The search returned nothing useful, and the LLM had no way to recover.

The flow that broke:

flowchart TD
    U([User: How do I run Olama?])
    S[search - Olama - no useful results]
    A([LLM: I don't have information about Olama.])

    U --> S --> A

The pipeline is fixed: search, build prompt, LLM.

def rag(self, query):
    search_results = self.search(query)
    prompt = self.build_prompt(query, search_results)
    answer = self.llm(prompt)
    return answer

The LLM is a passenger here, not a driver. It never sees the bad search results, so it can’t try again with a corrected query.

The agent alternative

An agent puts the LLM in charge.

Instead of running search ourselves, we give the LLM a search tool. It decides when to call it and what to search for.

The same typo question now goes like this:

flowchart TD
    U([User: How do I run Olama?])
    L1[LLM: I'll search for 'Olama']
    S1[search - Olama - no useful results]
    L2[LLM: Hmm, no results. Maybe a typo for 'Ollama'?]
    S2[search - Ollama - found results!]
    A([LLM: Here's how to run Ollama locally...])

    U --> L1 --> S1 --> L2 --> S2 --> A

The LLM searched, saw the results were bad, and decided to try again with a different query. It made that decision on its own. We didn’t write any code to handle typos.

The difference is about who makes the decisions:

The mechanism that makes this possible is function calling, and that’s what the rest of this lesson is about.

Asking without tools

First, let’s see what the LLM does without any tools. We ask it a course-specific question and look at the answer.

messages = [
    {"role": "user", "content": "I just discovered the course. Can I join it?"}
]
response = openai_client.responses.create(
    model="gpt-5.4-mini",
    input=messages,
)
response.output_text

The model answers from its general knowledge, something like “it depends on the course” or “check the course website”. It doesn’t know about our FAQ, so the answer is vague and not helpful. This is exactly why we need RAG, and why we want to hand the model a tool.

Defining the tool

First we define a top-level search function that queries the index directly. The model will reference it by this name. We keep the Python function and the tool name aligned so the dispatch is easier later.

def search(query):
    boost_dict = {"question": 3.0, "section": 0.5}
    filter_dict = {"course": "llm-zoomcamp"}

    return index.search(
        query,
        num_results=5,
        boost_dict=boost_dict,
        filter_dict=filter_dict
    )

Next we tell the model about this function. The model doesn’t see our Python code, only a schema describing what the function does and what arguments it takes. LLMs are language agnostic. At the end we’re just making an HTTP call, so we describe the tool in JSON rather than in Python. The same schema would work from TypeScript or Java.

search_tool = {
    "type": "function",
    "name": "search",
    "description": "Search the FAQ database for entries matching the given query.",
    "parameters": {
        "type": "object",
        "properties": {
            "query": {
                "type": "string",
                "description": "Search query text to look up in the course FAQ."
            }
        },
        "required": ["query"],
        "additionalProperties": False
    }
}

The description is the most important field, because the model reads it to decide when to call the functio

Key concepts

Sources

DataTalks.Club. Hosted on GitHub Pages. Built with Rustkyll. We use cookies.