ai-agent-book 精选快照（<2MB 代码与文档，来自 github.com/bojieli/ai-agent-book）

chapter2/prompt-engineering/tau_bench/envs/retail/data/__init__.py

@@ -0,0 +1,21 @@
+# Copyright Sierra
+
+import json
+import os
+from typing import Any
+
+FOLDER_PATH = os.path.dirname(__file__)
+
+
+def load_data() -> dict[str, Any]:
+    with open(os.path.join(FOLDER_PATH, "orders.json")) as f:
+        order_data = json.load(f)
+    with open(os.path.join(FOLDER_PATH, "products.json")) as f:
+        product_data = json.load(f)
+    with open(os.path.join(FOLDER_PATH, "users.json")) as f:
+        user_data = json.load(f)
+    return {
+        "orders": order_data,
+        "products": product_data,
+        "users": user_data,
+    }

chapter2/prompt-engineering/tau_bench/envs/retail/data/readme.md

@@ -0,0 +1,21 @@
+# Mock Data Generation
+
+## Current Mock Data for the Benchmark
+Feel free to use some of the data for other purposes.
+- `users.json`: a database of users with their emails, addresses, and orders
+- `products.json`: a database of products, where each product has variants (e.g., size, color).
+- `orders.json`: a database of orders that can be operated upon.
+
+
+Check `../tools` for mock APIs on top of current mock data.
+
+
+### Experience of Mock Data Generation
+
+Read our paper to learn more about the generation process for each database. In general, it involves the following stages:
+
+1. Design the type and schema of each database. Can use GPT for co-brainstorming but has to be human decided as it is the foundation of everything else.
+2. For each schema, figure out which parts can be programmaticly generated and which parts need GPT. For example,
+    - Product types (shirt, lamp, pen) and user names (Sara, John, Noah) need GPT generation
+    - Product price and shipping date can be generated via code
+3. Use GPT to generate seed data (first names, last names, addresses, cities, etc.), then use a program to compose them with other code generated data. Can use GPT to help write the code for this part, but I think code-based database construction is more reliable than GPT-based database construction (e.g., give some example user profiles and ask GPT to generate more --- issues with diversity and reliability).