traceverse-community/food-delivery-evals

Food Delivery — SmolTrace Tasks

Part of the TraceVerse Community open evaluation & observability stack.

A SmolTrace-format evaluation suite of 91 tasks for agents calling food delivery MCP servers, retargeted onto Swiggy's real published MCP surface.

---

Surface change — 2026-07-26

This dataset was originally written against an 18-tool mock surface (7 food / 6 instamart / 5 dineout) implemented by the companion *-mcp Spaces. It has been retargeted onto Swiggy's real published MCP surface — 35 tools across three servers (14 food / 13 instamart / 8 dineout), 109 parameters, 64 required.

Prompts are unchanged. They were written intent-shaped ("order biryani") rather than against fixed signatures, so they survived the surface change; only expected_tool and id were retargeted.

What was remapped

Old (mock) tool	Real tool	Rows
search_restaurants	food.search_restaurants	26
view_menu	food.get_restaurant_menu	19
place_order	food.place_food_order	20
track_delivery	food.track_food_order	13
apply_promo	food.apply_food_coupon	13

What was dropped

20 of 111 tasks exercised capabilities that do not exist anywhere on the real surface. They were not forced onto an approximately-similar tool — SmolTrace scores expected_tool by exact name membership, so a near-miss target would grade wrong behaviour as correct. They are preserved with their reason in dropped.jsonl.

Old tool	Rows	Why no real tool can carry it
rate_restaurant	10	No rating or review write API exists on any of the three servers. The real Food surface is discovery, cart, order, track and support only; nothing accepts a star rating or review text.
contact_driver	10	No driver-contact channel exists on any server. track_food_order returns delivery progress, but the published schema exposes no phone number, message endpoint or delivery-partner identifier to contact.

Also changed

• id values now carry the real tool name (food-delivery_search_restaurants_batch0_0). The batch/index numbers are preserved, and remap/audit/row-crosswalk.csv in the source repo maps every old id to its new one. Results from before 2026-07-26 are not comparable and should not be joined.
• The prompt template below was rewritten. The previous version demonstrated mock signatures that do not exist on the real surface; it was teaching candidate models a fictional API.

---

Dataset info

• Tasks: 91
• Format: SmolTrace evaluation format (6 columns, unchanged)
• Server: food — POST mcp.swiggy.com/food
• Tools covered: apply_food_coupon, get_restaurant_menu, place_food_order, search_restaurants, track_food_order
• License: Apache 2.0

Usage with SmolTrace

from datasets import load_dataset

dataset = load_dataset("traceverse-community/food-delivery-evals")
# smoltrace-eval --model openai/gpt-4o-mini --dataset-name traceverse-community/food-delivery-evals

Schema

Column	Type	Description
id	string	food-delivery_<tool>[_<suffix>]_batch<n>_<m>
prompt	string	Task description
expected_tool	string	One of: apply_food_coupon, get_restaurant_menu, place_food_order, search_restaurants, track_food_order
difficulty	string	easy / medium / hard
agent_type	string	tool (ToolCallingAgent) or code (CodeAgent)
expected_keywords	list[string]	2-7 keywords expected in plan/output

How expected_tool is scored

SmolTrace counts membership in the trajectory (tools_used.count(expected_tool) > 0), not the first call. So expected_tool names the goal tool: a task mapped to apply_food_coupon may legitimately require several discovery calls first. Note also that for agent_type: "code" rows, SmolTrace computes correct_tool but does not require it for success — those rows score on response keywords alone.

Example row

{
    "id": "food-delivery_search_restaurants_batch0_0",
    "prompt": "I'm craving Italian food. Can you find some Italian restaurants near me?",
    "expected_tool": "search_restaurants",
    "difficulty": "easy",
    "agent_type": "tool",
    "expected_keywords": [
        "Italian",
        "restaurants",
        "search"
    ]
}

Distribution

• By tool: search_restaurants 26 · place_food_order 20 · get_restaurant_menu 19 · track_food_order 13 · apply_food_coupon 13
• By difficulty: easy 60 · medium 17 · hard 14
• By agent type: tool 68 · code 23

Known limitations

• Fixture-bound prompts. Some prompts name entities invented for the mock world (order ids like #12340, product ids like p-204, mock promo codes) or quote USD amounts where Swiggy operates in INR. Their expected_tool is correct, but they may be unanswerable against real fixture data. See remap/audit/fixture-bound.csv.
• Capability drift in prompts. Some prompts assume filters the real surface does not expose (minimum rating, cost-for-two, gluten-free). The intent still resolves to the right tool, but the model cannot satisfy the filter through arguments.
• Weak keyword grading. expected_keywords passes if any one keyword appears in the response (case-insensitive substring). Many rows include generic terms, so response_correct is close to always-true. Treat correct_tool as the meaningful signal here.

---

Prompt template

Save the YAML below as prompt_template.yaml and pass via system_prompt_path to smolagents.ToolCallingAgent or CodeAgent. Argument names below are validated against the published schema at generation time.

system_prompt: |-
  You are an expert assistant for Swiggy Food Delivery. You call the real Swiggy MCP
  tools listed below. Use ONLY the parameters each tool actually declares.

  Each step is Action (tool call) → Observation. Finish with `final_answer`.

  Example:

  Task: "Order a Margherita pizza from a top-rated Italian place."
  Action:
  {"name": "get_addresses", "arguments": {}}
  Observation: [{"id": "addr_01HXYZ", "addressLine": "221B Indiranagar", ...}]
  Action:
  {"name": "search_restaurants", "arguments": {"addressId": "addr_01HXYZ", "query": "Italian"}}
  Observation: {"restaurants": [{"id": "rest_42", "name": "...", "availabilityStatus": "OPEN", "distanceKm": 2.1}]}
  Action:
  {"name": "search_menu", "arguments": {"addressId": "addr_01HXYZ", "query": "Margherita pizza", "restaurantIdOfAddedItem": "rest_42"}}
  Observation: {"items": [{"menu_item_id": "itm_9", "name": "Margherita", "hasAddons": false}]}
  Action:
  {"name": "update_food_cart", "arguments": {"restaurantId": "rest_42", "addressId": "addr_01HXYZ", "cartItems": "[...]"}}
  Observation: {"success": true}
  Action:
  {"name": "get_food_cart", "arguments": {"addressId": "addr_01HXYZ"}}
  Observation: {"total": 349, "availablePaymentMethods": ["UPI"]}
  Action:
  {"name": "place_food_order", "arguments": {"addressId": "addr_01HXYZ"}}
  Observation: {"orderId": "ord_77"}
  Action:
  {"name": "final_answer", "arguments": {"answer": "Swiggy order placed successfully. Order ord_77, Margherita from rest_42."}}

  Available tools:
  {%- for tool in tools.values() %}
  - {{ tool.to_tool_calling_prompt() }}
  {%- endfor %}

  Rules: always tool-call; pass values not variables; never repeat with identical args; finish with final_answer.

planning:
  initial_plan: |-
    Survey known facts, list facts to look up, then write a step-by-step plan ending with <end_plan>.
    Task: {{task}}

The example above shows the full happy path. Note that Swiggy's published agent_guidance for get_addresses says to stop and let the user choose an address before calling anything else; in an autonomous eval there is no user, so the template proceeds with the first address. That is a deliberate deviation from Swiggy's guidance, made explicit rather than hidden.

---

Reproducing this dataset

hf clone datasets/traceverse-community/food-delivery-evals
cd food-delivery-evals
python3 generate.py > tasks.jsonl

The generator emits the remapped rows directly; dropped.jsonl is an audit artefact and is not regenerated by it.

---

Part of the TraceVerse Community stack

Project	Where
Trace viewer	TraceMind
Instrumentation	genai-otel-instrument
Evaluation framework	SmolTrace

Sibling datasets:

• instamart-evals
• dineout-evals

Citation

@misc{traceverse-food-delivery-evals,
  author    = {TraceVerse Community},
  title     = {Food Delivery Evals: a SmolTrace evaluation dataset for Swiggy MCP agents},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/traceverse-community/food-delivery-evals}
}

Acknowledgement

The SmolTrace evaluation format and prompt-template structure originated in the work of MCP-1st-Birthday on the TraceMind MCP Server's synthetic data generation. Tool schemas are derived from Swiggy's published MCP builder documentation at https://mcp.swiggy.com/builders/docs/reference/.