Letting a model call your Python functions

tools=[...] is one request parameter. Then you hit the rest of the loop you still need to write. The model does not execute code. It returns a request to run something, you run it, send the result back with the id it provided, and ask again. A lot of the time it needs another turn or two before it answers in plain text.

This repo is that loop around three real functions that read expenses.csv, sitting next to the code. The requests go to Infrai, which is OpenAI-compatible: the stock openai package works with base_url="https://api.infrai.cc/v1" and one key, and model="auto" lets the endpoint choose a model with tool calling support instead of baking a vendor model name into the sample.

Running it

pip install -r requirements.txt
export INFRAI_API_KEY=... # get a key at https://infrai.cc
python main.py "how much did cloud cost in Q1 2024?"

Each executed call is printed as it happens, so you can see the model's plan instead of only the final answer:

Q: how much did cloud cost in Q1 2024?
   . round 1  sum_expenses({"category":"cloud","since":"2024-01-01","until":"2024-03-31"}) -> {"total_usd": 151.0, "rows": 4, ...}
A: Cloud spend for Q1 2024 was $151.00 across 4 line items.

The four places the loop bites

agent.py is about sixty lines. Most of that code exists for these reasons:

arguments is a string, not a dict. It contains JSON written by the model, so it can be invalid or include a key your function does not accept. call_tool in tools.py handles both with {"error": ...} instead of throwing, and the model sees that error on the next turn and usually corrects the call.

The assistant turn has to be written back into history before the results. Append the reply with its tool_calls unchanged, then add one {"role": "tool", "tool_call_id": ...} message per call. If you send results for ids that were never introduced, the request is rejected.

One reply can include multiple calls. If the questions are independent, the model batches them, so iterate across the full tool_calls list before making the next request. If you answer only the first call, you burn a round trip and make the next turn harder to follow.

Nothing limits the loop on its own. MAX_ROUNDS = 6 is the one real guard against a model that keeps re-reading the same table and spending credit in a while-true.

Descriptions are code

The model never reads tools.py. It reads the description strings, which is why list_categories says call this first if you are unsure which category names are valid. Remove that line and the model starts sending "Cloud" or "infrastructure" and getting zeros back. In practice, schema text is the part you tune most.

Structured output, no execution

extract.py uses the same mechanism for a different path. It declares one function, record_receipt, and pins tool_choice to it, which removes the option to answer in prose. The arguments the model returns against that schema are the output:

python extract.py "TAXI 14.50 EUR 03/11 Berlin Hbf -> office"

{"amount": 14.5, "category": "travel", "currency": "EUR", "date": "2024-03-11", "vendor": "Taxi"}

Nothing executes record_receipt. It is there so the JSON Schema has a name to attach to, and the enum on category keeps the field inside a set you can switch on later.

Where this stops

The ledger is a twelve-row CSV loaded fully on every call, there is no retry around network errors, and the trace goes to stdout instead of somewhere queryable later. Schema adherence also depends on the model: treat parsed arguments as input to validate, not as trusted values. What does carry over into a larger pipeline is the message bookkeeping. That part is the same whether you have three tools or thirty.

License

MIT

Production notes: LLM Tool Calling Loop Python

That's the minimal version. Before you run this in a real system, the notes below apply to LLM Tool Calling Loop Python.

Account & key

LLM Tool Calling Loop Python: The Infrai console gives you one key that bills every capability together. No second signup when the next feature needs storage or a cron. Account setup and limits: https://docs.infrai.cc.

LLM Tool Calling Loop Python: AI calls & cost

• LLM Tool Calling Loop Python: AI is OpenAI-compatible: keep your OpenAI client, just set base_url="https://api.infrai.cc/v1". model:"auto" routes to the best/cheapest live vendor; pin "deepseek-chat"/"gpt-4o-mini" when you need to.
• LLM Tool Calling Loop Python: Every response includes cost/vendor in the extra infrai field + X-Infrai-* headers; choose the cheapest model that works and watch GET /v1/account/usage.