lib/utils.py

import os
from pathlib import Path

import openai
import tiktoken
import pandas as pd
from openai.embeddings_utils import get_embedding, cosine_similarity

encoding_name = "p50k_base"

encoding = tiktoken.get_encoding(encoding_name)

embedding_model = "text-embedding-ada-002"
openai.api_key = os.environ.get("OPENAI_API_KEY", None)

# read from current directory
df = pd.read_pickle(Path(__file__).resolve().parent.__str__() + "/../data/lethain.pkl")


def search_reviews(df, query):
    query_embedding = get_embedding(
        query,
        engine="text-embedding-ada-002"
    )
    df["similarity"] = df.embeddings.apply(lambda x: cosine_similarity(x, query_embedding))

    results = (
        df.sort_values("similarity", ascending=False)

    )
    return results


def construct_prompt(question: str, df: pd.DataFrame) -> str:
    MAX_SECTION_LEN = 500
    SEPARATOR = "\n* "

    separator_len = len(encoding.encode(SEPARATOR))

    f"Context separator contains {separator_len} tokens"

    """
    Fetch relevant
    """
    result = search_reviews(df, question)

    chosen_sections = []
    chosen_sections_len = 0
    chosen_sections_indexes = []

    for section_index, row in result.iterrows():
        # Add contexts until we run out of space.

        tokens_num = len(encoding.encode(row.content))
        chosen_sections_len += tokens_num
        if chosen_sections_len > MAX_SECTION_LEN:
            break

        chosen_sections.append(SEPARATOR + row.content.replace("\n", " "))
        chosen_sections_indexes.append(str(section_index))

    # Useful diagnostic information
    print(f"Selected {len(chosen_sections)} document sections:")
    print("\n".join(chosen_sections_indexes))

    header = """You name is Will Larson, you are CTO at Calm and a blogger about engineering leadership. Answer the question as truthfully as possible using the provided context, and if the answer is not contained within the text below, say "I don't know."\n\nContext:\n"""

    return header + "".join(chosen_sections) + "\n\n Q: " + question + "\n A:"


def ask(question):
    prompt = construct_prompt(question, df)
    result = openai.Completion.create(
        prompt=prompt,
        temperature=0,
        max_tokens=300,
        model="text-davinci-003"
    )

    return result['choices'][0]['text']