Exercise 2: Multi-Tool Selection¶

Corresponds to Stage 3 — Tool Use & Agent Intro Exercise 2.

Why this matters¶

This exercise puts an LLM in front of three tools in a single turn: web_search, calculator, calendar_lookup. The point isn't tool quality — it's watching how schema name / description / parameters steer the model's choice. Writing schemas well is one of the highest-leverage things you do in Stage 3.

How to run — two paths¶

Path A (default, free, local)¶

pip install -r requirements.txt
ollama pull qwen2.5:3b
ollama serve
python starter.py

Budget: $0. A single qwen2.5:3b tool call takes ~1-5s (CPU slower, GPU faster).

Path B (Anthropic, cloud-quality comparison)¶

pip install -r requirements.txt
export ANTHROPIC_API_KEY=sk-ant-...
python starter_anthropic.py

Budget: ~$0.0005 per run (claude-haiku-4-5).

Expected output (Path A, local):

❓ Question: What is (19 * 42) - 8? Use the best available tool. (using Ollama qwen2.5:3b)
   tool: calculator
   tool_input: {'expression': '(19 * 42) - 8'}
   observation: 790
✅ Exercise 2 passed — you ran multi-tool selection locally on qwen2.5:3b, $0/run

Validate the logic without API credits (mock-based)¶

python test.py            # validates Path A (Ollama) starter.py logic
python test_anthropic.py  # validates Path B (Anthropic) starter_anthropic.py logic

Both test suites use unittest.mock, no real API call, $0/run. Path A uses the OpenAI-compat response shape; Path B uses Anthropic content blocks.

SDK differences between the two paths¶

Three key differences (everything else is identical):

Part	Anthropic (Path B)	OpenAI-compat / Ollama (Path A)
Schema wrap	`tools=[{name, description, input_schema}, ...]`	`tools=[{"type": "function", "function": {name, description, parameters}}, ...]`
Reading tool call	`resp.content[i].type == "tool_use"`	`resp.choices[0].message.tool_calls[i]`
input format	`call.input` is already a dict	`call.function.arguments` is a JSON string — needs `json.loads(...)`

The selection logic is backend-agnostic — write a good schema and qwen2.5:3b picks the right tool too. This exercise is a great place to compare "on which questions does Claude pick the right tool but qwen2.5 doesn't?" — a clean way to feel the boundary of small models.

Common pitfalls¶

The most common failure in multi-tool design is descriptions that read like documentation, not decision rules:

calendar_lookup described as "calendar" is ambiguous with web_search; "look up events for a specific date" is clearer
web_search is for "external / recent / uncertain info", calculator for arithmetic — the clearer the boundary, the fewer wrong picks
Small models (qwen2.5:3b) are more sensitive to description quality than Claude — the same schema where Claude might guess correctly can lead qwen astray

Want smarter answers?¶

Default is claude-haiku-4-5 (cheapest). Try Sonnet:

MODEL=claude-sonnet-4-6 python starter_anthropic.py

Or on the Ollama path, swap to qwen2.5:7b (bigger, more stable, but slower):

MODEL=qwen2.5:7b python starter.py

Extensions¶

Add more tools — append one entry each to TOOLS_SPEC + TOOL_IMPL
Make it multi-turn ReAct — wrap the single call in a while loop; see ../03-react-from-scratch/
Dig into schema design — see ../06-schema-design/ for a bad vs good schema A/B