few_shot_prompt_llama.py

from langchain import OpenAI, LLMChain, PromptTemplate
from langchain import PromptTemplate, FewShotPromptTemplate
from langchain.chains import LLMChain
from langchain.llms import LlamaCpp
from langchain.prompts import PromptTemplate
import csv

examples = [
    {"product": "Toothpaste",
     "category": "Health:Dental"},
    {"product": "Toilet Flapper",
     "category": "Home:Maintenance"},
    {"product": "Laptop Stand",
     "category": "Computer:Accessories"},
    {"product": "Pressure Cooker",
     "category": "Kitchen:Appliances"},
    {"product": "T-shirt",
     "category": "Clothing"},
    {"product": "Bananas",
     "category": "Grocery"},
]

llm = LlamaCpp(model_path="../llama.cpp/models/ggml-alpaca-7b-q4.bin")

example_formatter_template = """
Product: {product}
Category: {category}\n
"""
example_prompt = PromptTemplate(
    input_variables=["product", "category"],
    template=example_formatter_template,
)

few_shot_prompt = FewShotPromptTemplate(
    examples=examples,
    example_prompt=example_prompt,
    prefix="Give the category of every product",
    suffix="Product: {product}\nCategory:",
    input_variables=["product"],
    example_separator="\n\n",
)

llm_chain = LLMChain(
    llm=llm,
    prompt=few_shot_prompt,
    verbose=True,
)

with open('../../my/finances/amz.csv', 'r') as file:
    reader = csv.DictReader(file)
    for row in reader:
        print(llm_chain.predict(product=row['Product Name']))

This takes an Amazon order history CSV and categorizes products for expense tracking software such as Beancount. An example run produced:

Product: LEVN Bluetooth Headset with Microphone, Trucker Bluetooth Headset with AI Noise Cancelling & Mute Button, Wireless On-Ear Headphones 60 Hrs Working Ti
Category:

llama_print_timings:        load time =   534.18 ms
llama_print_timings:      sample time =    10.94 ms /    23 runs   (    0.48 ms per run)
llama_print_timings: prompt eval time = 15154.56 ms /   167 tokens (   90.75 ms per token)
llama_print_timings:        eval time =  5072.16 ms /    22 runs   (  230.55 ms per run)
llama_print_timings:       total time = 20240.41 ms

> Finished chain.
 Electronics:Headsets