Designing multi-step AI workflows

Introduction to Generative AI in Snowflake

James Cha-Earley

Senior Developer Advocate, Snowflake

Cortex review workflow

Goal: Build an automated system for international guest feedback

Translate each review to English
Summarize the main point
Categorize it for routing to the right team
Generate a response
Translate it back to the original language

Workflow where a review is received, translated, summarized and categorized, and a response is sent

Extracting reviews

-- SQL cell
SELECT DESCRIPTION
FROM HOTELS.REVIEWS
WHERE LANGUAGE = 'es'
LIMIT 1;

# Python cell
df = cell1.to_pandas()
review_text = df["DECRIPTION"].iloc[0]

Spanish review

print(review_text)

Buen hotel y bien situado pero desafortunadamente no me toco buena suerte con 
el servicio; el aire no servía; pedí la cama extra tres veces; la cafetera estaba
dañada; y me sacaron las maletas fuera porque argumentaron que no hice Check out;
siendo que era un día despues; pero su sistema lo marco antes; una total
descortesía hecharme del cuarto y cancelar mis llaves; regresar de caminar y darse
cuenta que han tomado tus cosas fuera es increíble y mas aun sin ninguna disculpa;
yo no volvería ahí aunque la vrd el hotel es bueno y su jubilación el trato
me decepciono

Translation

translated = translate(
    text=review_text,
    from_language="es",
    to_language="en"
)

print(translated)

Good hotel and well located, but unfortunately, I didn't have good luck with the service; the air
conditioning didn't work; I asked for the extra bed three times; the coffee maker was broken; 
and they took my bags out because they argued that I hadn't checked out; even though it was a day
later; but their system marked it as checked out; a total discourtesy to throw me out of the room
and cancel my keys; returning from a walk and realizing they've taken your things out is
incredible and even more so without any apology; I wouldn't go back there even though the hotel
itself is good and their retirement the treatment disappoints me.

Summarization

summary = summarize(text=translated)

print(summary)

The hotel was well-located, but the service was disappointing. The air conditioning
didn't work, an extra bed was not provided despite multiple requests, and the coffee
maker was broken. The hotel staff took the complainant's bags and canceled their
keys despite a later checkout date, which was discourteous and left the complainant
feeling disappointed.

Classification

topic = classify_text(
    text=summary,
    labels=["staff", "cleanliness", "pricing", "room", "food"]
)

print(topic)

{
  "label": "staff"
}

Text generation

response = complete(
        prompt=f"Write a brief and professional response to this review: {summary}",
        model='llama3.1-8b',
        options={'temperature':0.3, 'max_tokens':120})

print(response)

Thank you for sharing your feedback. While we're glad you found the location
convenient, we're truly sorry to hear about the service issues you experienced.
We understand how frustrating it must have been to face multiple inconveniences
during your stay. Your comments have been shared with the team to ensure these concerns
are addressed and do not recur.

Response translation

translated_response = translate(
    text=response,
    from_language="en",
    to_language="es"
)

print(translated_response)

Gracias por compartir sus comentarios. Si bien nos alegra saber que encontró conveniente
la ubicación, lamentamos sinceramente los inconvenientes que experimentó con el servicio.
Entendemos lo frustrante que debió haber sido enfrentar múltiples inconvenientes durante
su estadía. Sus comentarios han sido compartidos con el equipo para asegurarnos de que
estas situaciones se aborden y no vuelvan a ocurrir.

Cortex cost model

A cortex workflow showing that input tokens, compute, and output tokens all have an associated cost

¹ Image generated by ChatGPT-4o

Limiting cost

Trim inputs, process only relevant text
Limit output size
Lower temperature

# Limit cost of complete
complete(prompt=prompt, 
         model='llama3.1-8b', 
         options={
            'max_tokens':120,

            'temperature':0})

Cortex best practices

Effectively chain models

# Summarize first if calling multiple downstream functions
summarize()
text_classify()
complete()
translate()

Logging
Caching
Batch pipelines

Let's practice!

Introduction to Generative AI in Snowflake