Augmenting prompts with Python

Web scraping

Fetch rich prompt material with Python

Using an http client to fetch HTML that an LLM will filter for a list[RelatedArticle]:

import bs4
import httpx
import marvin
from typing_extensions import TypedDict

class RelatedArticle(TypedDict):
    title: str
    link: str


@marvin.fn
def retrieve_HN_articles(topic: str | None = None) -> list[RelatedArticle]:
    """Retrieve only articles from HN that are related to a given topic"""
    response = httpx.get("https://news.ycombinator.com/")
    soup = bs4.BeautifulSoup(response.text, 'html.parser')
    return [
        (link.text, link['href']) for link in soup.select('.titleline a')
    ]

retrieve_HN_articles("rust")

Result

[
    {
        'title': 'A lowering strategy for control effects in Rust',
        'link': 'https://www.abubalay.com/blog/2024/01/14/rust-effect-lowering'
    },
    {
        'title': 'Show HN: A minimal working Rust / SDL2 / WASM browser game',
        'link': 'https://github.com/awwsmm/hello-rust-sdl2-wasm'
    }
]

Note

You could also use marvin.extract to extract the list[RelatedArticle] from the output of the un-decorated function retrieve_HN_articles:

related_articles = marvin.extract(retrieve_HN_articles(), RelatedArticle)

Vectorstore-based RAG

Stuff top k document excerpts into a prompt

Using an http client to fetch HTML that an LLM will filter for a list[RelatedArticle]:

from typing_extensions import TypedDict
import marvin
from marvin.tools.chroma import query_chroma # you must have a vectorstore with embedded documents
from marvin.utilities.asyncio import run_sync

class Answer(TypedDict):
    answer: str
    supporting_links: list[str] | None

@marvin.fn
def answer_question(
    question: str,
    top_k: int = 2,
    style: str = "concise"
) -> Answer:
    """Answer a question given supporting context in the requested style"""
    return run_sync(query_chroma(question, n_results=top_k))

answer_question("What are prefect blocks?", style="pirate")

Result

{
    'answer': "Ahoy! Prefect blocks be a primitive within Prefect fer storin' configuration and interfacin' with th' external systems. Ye can use 'em to manage credentials and interact with services like AWS, GitHub, and Slack. Arr, they be comin' with methods for uploadin' or downloadin' data, among other actions, and ye can register new ones with Prefect Cloud or server.",
    'supporting_links': ['https://docs.prefect.io/latest/concepts/blocks/']
}