S116 · Prompting — Zero-Shot, Few-Shot, Chain-of-Thought, ReAct

"""prompt_patterns.py — measure zero/few/CoT/ReAct on a small task set.
 
Requires: openai or anthropic sdk + API key. Change MODEL/CLIENT to match.
"""
from __future__ import annotations
import json
import os
import re
import time
from openai import OpenAI
 
client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
MODEL = "gpt-4o-mini"     # cheap, fast, plenty smart enough for the demo
 
 
def ask(system: str, user: str, temperature: float = 0.0) -> str:
    r = client.chat.completions.create(
        model=MODEL,
        messages=[{"role": "system", "content": system},
                  {"role": "user",   "content": user}],
        temperature=temperature,
    )
    return r.choices[0].message.content.strip()
 
 
# ---------- The evaluation set: a math word problem + an extraction ----------
MATH_Q = (
    "Alice has 3 more apples than Bob. Bob has twice as many as Charlie. "
    "Together they have 33 apples. How many does Charlie have?"
)
MATH_A = "6"
 
EXTRACT_Q = (
    "Customer email: 'Hi, my name is Priya Sharma. My order #A-4821 arrived "
    "damaged. Please refund $89.99 to my card ending 4242. Thanks!'"
)
EXTRACT_SCHEMA = ("Extract as JSON: {name, order_id, amount, "
                  "issue_category (damage/late/wrong/other)}")
 
 
# ---------- 1. Zero-shot ----------
def zero_shot(q: str) -> str:
    return ask("You are a helpful assistant. Answer concisely.", q)
 
# ---------- 2. Few-shot ----------
FEW_SHOT_MATH = """Solve each problem. Give only the final integer answer.
 
Q: There are 5 pens. I buy 2 boxes of 4. How many pens total?
A: 13
 
Q: A train leaves at 3 and arrives at 7. How many hours?
A: 4
 
Q: {q}
A:"""
def few_shot(q: str) -> str:
    return ask("You are a helpful assistant.", FEW_SHOT_MATH.format(q=q))
 
# ---------- 3. Chain-of-thought ----------
COT = "Let's think step by step. Show your reasoning, then on the last line write 'ANSWER: <value>'."
def cot(q: str) -> str:
    out = ask("You are a helpful assistant.", f"{q}\n\n{COT}")
    m = re.search(r"ANSWER:\s*(.+)", out)
    return m.group(1).strip() if m else out.splitlines()[-1]
 
# ---------- 4. ReAct (tool-using) — mini implementation ----------
def get_apples(person: str) -> int:  # a fake 'database'
    return {"charlie": 6, "bob": 12, "alice": 15}.get(person.lower(), -1)
 
TOOLS_DESC = ("You have one tool: get_apples(person). "
              "Respond in the form:\n"
              "Thought: ...\nAction: get_apples('name')\n"
              "or\nThought: ...\nFinal Answer: <value>")
def react(q: str, max_steps: int = 5) -> str:
    convo = f"Question: {q}\n"
    for _ in range(max_steps):
        out = ask(TOOLS_DESC, convo)
        convo += out + "\n"
        if "Final Answer:" in out:
            return out.split("Final Answer:")[-1].strip()
        m = re.search(r"Action:\s*get_apples\(['\"]?(\w+)['\"]?\)", out)
        if m:
            obs = get_apples(m.group(1))
            convo += f"Observation: {obs}\n"
        else:
            break
    return "(no final answer)"
 
 
# ---------- Run all four on the MATH problem ----------
print(f"MATH problem: {MATH_Q}")
print(f"Correct answer: {MATH_A}\n")
for name, fn in [("zero_shot", zero_shot),
                 ("few_shot",  few_shot),
                 ("cot",       cot),
                 ("react",     react)]:
    t0 = time.time()
    ans = fn(MATH_Q)
    print(f"  {name:<10} → {ans[:60]:<60}  ({time.time()-t0:.1f}s)")
 
# ---------- Run zero-shot vs few-shot on EXTRACTION ----------
print(f"\nEXTRACTION problem: {EXTRACT_Q}")
zs = ask("You are a data extractor. Respond with ONLY valid JSON.",
         f"{EXTRACT_Q}\n\n{EXTRACT_SCHEMA}")
print(f"  zero-shot JSON → {zs[:120]}")