How to Engineer Industrial‑Scale Function Call Data for AI Agents

Function‑call agents face several practical challenges: low coverage of handcrafted data, fragile JSON formats, and the inability to learn procedural flows. To build production‑grade agents, the training data must be highly engineered and reproducible.

Why a Data Sandbox Is Essential

Construct a controllable, extensible, and reproducible "data sandbox".

The sandbox defines every variable, label, and workflow at the data level, then automatically generates a complete, multi‑turn conversation dataset.

Three Perspectives Covered

Why function‑call data needs a sandbox

The complete label system and variables

From sandbox to real dialogues

1) Interview Question: Why Not Hand‑Write SFT Data?

Hand‑written examples suffer from:

Coverage too low – even hundreds of examples cannot span all real scenarios.

Format errors – missing JSON encoding, wrong tool_calls structure, role switches, mismatched tool_call_id, etc.

Models never learn the underlying process, only isolated snippets.

2) Data Sandbox Architecture

Using a travel‑assistant agent as an example, the sandbox contains four core workflows:

Travel planning (RAG + weather)

Navigation (map)

Hotel search (recommendation + reviews)

Chit‑chat / rejection

The sandbox must enumerate all mutable variables across these workflows.

Step 1: Define the Tag System (the soul of function calls)

Travel planning – no follow‑up (320 examples)

Travel planning – needs follow‑up (40)

Navigation – no follow‑up (80)

Navigation – needs follow‑up (16)

Hotel query – no follow‑up (160)

Hotel query – needs follow‑up (32)

Travel‑related chit‑chat (80)

Non‑travel rejection (80)

Each tag guarantees controllability and sufficient coverage.

Step 2: Define Business Variables (user profile)

Every data point includes a system prompt with user information such as name, city ID, departure date, and starting coordinates:

{
  "role": "system",
  "content": "## User Info
- Name: 吴师兄
- City ID: 1012510801
- Departure Date: 2025-12-02
- Start Coord: 100.479921,59.1237401"
}

Step 3: Query Templates (semantic perturbations)

For hotel search, 30 different phrasings are generated, e.g.:

帮我找一家上海外滩附近 2000 左右的酒店
外滩景观好的酒店推荐
我周五去上海，住两晚，预算 2000，能推荐吗
在魔都哪里住方便一点？
…

Step 4: Tool Result Variations

Tool returns empty

Returns 1 hotel

Returns 3 hotels

Returns 10 hotels

Returns invalid review

Returns error

Step 5: Dialogue Branches (follow‑up vs direct)

Missing destination → must ask

Missing dates → must ask

Complete info → direct tool chain

Data Generation Pipeline

Stage 1 – Seed Data Generation (generate_dataset.py)

class DatasetGenerator:
    def __init__(self):
        self.names = [...]
        self.cities = {...}
        self.hotel_queries = [...]
        self.travel_queries = [...]
        self.route_queries = [...]

    def generate_travel_plan_no_ask(self):
        return {
            "用户问题": random.choice(self.travel_queries),
            "用户名字": random.choice(self.names),
            "用户所处城市": city_id,
            "出发日期": date,
            "起点坐标": coord,
            "类型": "旅行规划-不需要反问",
            "是否追问": "否"
        }

Each label has a dedicated generator function, producing a total of 1,010 seed records.

Stage 2 – Convert Seeds to Full Dialogues

The core of a function‑call project is converting seed data into OpenAI‑compatible conversation JSON:

[
  {"role": "system", "content": "..."},
  {"role": "user", "content": "我要在武汉住酒店，预算200-300元"},
  {"role": "assistant", "tool_calls": [{"id": "call_9e45f8c7", "type": "function", "function": {"name": "recommend_hotels", "arguments": "{\"requirements\": \"武汉，预算200-300元\"}"}}]},
  {"role": "tool", "content": "...工具返回...", "tool_call_id": "call_9e45f8c7"},
  {"role": "assistant", "content": "已根据您的预算，为您找到以下酒店…"}
]

The generation algorithm follows these steps:

Select workflow based on label.

Determine whether a follow‑up is required.

Automatically generate the tool‑call chain.

Generate synthetic tool responses.

Generate the final natural‑language answer.

All steps are executed by code, eliminating manual dialogue writing.

What the Model Learns

Intent routing : Given "我想住两晚 2000 元内的酒店", the model selects the hotel‑search intent, detects missing dates, and follows the "need follow‑up" branch.

Multi‑turn follow‑up : The assistant asks for missing dates and processes the user’s answer.

Tool sequencing : For hotel search, the model calls recommend_hotels, then get_hotel_reviews, and finally produces the answer.

Handling empty tool results : The sandbox includes scenarios where a tool returns nothing, teaching the model to reply with a graceful apology.

Rejecting out‑of‑domain queries : Non‑travel questions trigger a domain‑specific refusal.

Ensuring Coverage and Quality

Because every variable and branch is defined in the sandbox, the dataset achieves:

High controllability – each conversation’s logic is explicit.

Full coverage – every label has enough examples.

Reproducibility – the same code can regenerate the entire set.

This engineered data dramatically improves the success rate of function‑call agents in real‑world deployments.