← All articles

6 min readAIMarketingGroqLPU

Cost-Effective AI Prompt Engineering for Small Business

Small businesses can now afford AI‑generated content, thanks to Groq’s sub‑$0.001 per 1k tokens, making on‑demand copy generation financially viable for boutique retailers.

The Cost Barrier and the LPU Advantage

AI‑generated marketing copy has become a staple in larger brands, but the upfront cost of large language models (LLMs) has kept many small shops on the sidelines. In 2025, AI spending rose 27 % year‑over‑year, yet surveys show that cost and data privacy remain the top obstacles for small‑business owners. Groq’s 2026 LPU roadmap changes that equation: a single LPU can process a thousand tokens for less than a cent, and the hardware’s low power profile means it can run on modest edge devices or in a lightweight cloud instance.

For a local furniture shop, a single promotional paragraph can cost a few cents in a traditional LLM setup. With Groq, the same paragraph can be generated for less than a cent, freeing the shop’s budget for inventory or local advertising. The key is to design prompts and workflows that keep token usage (and therefore cost) under tight limits while still producing brand‑consistent copy.

Selecting the Right Model Size

Groq offers a range of LPU models, from lightweight 4‑billion‑parameter engines to larger 16‑billion‑parameter variants. The choice depends on the complexity of the content and the need for nuance. For short‑form marketing (e.g., a 140‑character Instagram caption or a 30‑second radio script), the 4‑billion‑parameter LPU is more than adequate. It delivers natural language fluency, retains context over a few hundred tokens, and does so with minimal latency.

When the content requires deeper context, such as a product description that references multiple features, a 8‑billion‑parameter LPU can be justified. The extra capacity reduces hallucination risk and improves consistency across multiple outputs. However, the cost per token rises proportionally, so the decision should be guided by the value each extra nuance brings to the customer experience.

Crafting Prompt Templates that Respect Brand Voice

A prompt is the recipe that tells the LLM what to produce. For a local shop, the prompt must balance creativity with brand guidelines. A template that works well across several clients looks like this:

You are a copywriter for {shop_name}. Write a {length} promotional piece in {tone} that highlights {focus}. The copy should:
- Use {language_style} language
- Avoid jargon
- End with a call‑to‑action: {cta}

Replace the placeholders with data stored in the shop’s database. The prompt length is kept short (typically 30–40 tokens), so that the LLM’s output dominates the token budget. By standardizing the prompt structure, the edge function can generate multiple copies with a single call, and the LLM’s internal state is reused efficiently.

A practical example for a boutique bakery:

You are a copywriter for Sweet Crumbs Bakery. Write a 140‑character promotional piece in cheerful tone that highlights our new almond croissants. The copy should:
- Use playful language
- Avoid jargon
- End with a call‑to‑action: "Try them today!"

The output will be concise, brand‑aligned, and ready for posting.

Token Budgeting: Keeping Costs Low

The LLM’s cost is directly proportional to the number of tokens processed. For a 140‑character tweet, the output is roughly 25 tokens. Adding the 40‑token prompt brings the total to 65 tokens. At $0.000001 per token, that’s $0.000065 per copy, well below the $0.01 threshold.

A few tricks help keep the token count minimal:

  1. Pre‑trimmed Prompts – Keep the prompt to essential instructions. Avoid verbose explanations.
  2. Output Length Limits – Use the LLM’s max_tokens parameter to cap the response.
  3. Batching – When generating several copies in one request, the prompt can be shared across all, reducing repeated overhead.
  4. Post‑processing – Strip any unwanted formatting or placeholders before sending the final text to the shop’s social media scheduler.

By applying these measures, a local retailer can generate dozens of unique promos per day for pennies.

Edge‑First Integration with Supabase

Supabase Edge Functions provide a lightweight, serverless environment that runs close to the data source. In the Sottocasa.ai project, the edge function acts as a proxy between the mobile app and Groq’s API, ensuring that user data never leaves the edge unless explicitly allowed.

A simplified edge function looks like this:

export default async function handler(req: Request) {
  const { shopId, promptParams } = await req.json()

  // Retrieve brand settings from Supabase
  const { data: shop } = await supabase
    .from('shops')
    .select('name, tone, language_style, cta')
    .eq('id', shopId)
    .single()

  // Build the prompt
  const prompt = `
    You are a copywriter for ${shop.name}. Write a ${promptParams.length} promotional piece in ${shop.tone} that highlights ${promptParams.focus}. The copy should:
    - Use ${shop.language_style} language
    - Avoid jargon
    - End with a call‑to‑action: ${shop.cta}
  `

  // Call Groq
  const response = await fetch('https://api.groq.com/v1/chat/completions', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json', 'Authorization': `Bearer ${GROQ_API_KEY}` },
    body: JSON.stringify({
      model: 'groq-4b',
      messages: [{ role: 'user', content: prompt }],
      max_tokens: promptParams.maxTokens,
    }),
  })

  const json = await response.json()
  return new Response(JSON.stringify({ text: json.choices[0].message.content }))
}

The edge function keeps latency low (under 150 ms for a 4‑billion‑parameter LPU) and ensures that the only data leaving the shop’s environment is the prompt and the resulting text, which can be stored or forwarded to marketing channels.

Privacy‑First Data Flow: GDPR Compliance

Small businesses in the EU must adhere to GDPR, which restricts the transfer of personal data outside the European Economic Area (EEA). The edge‑first approach keeps raw customer data (names, purchase history, preferences) on the Supabase instance. Only non‑personal, brand‑related prompts are sent to Groq. This satisfies the “data minimisation” principle.

For added protection, the function can hash any identifiers before sending them to the LLM. The LLM then receives only the hash, which it can use to reference the original data internally, but cannot reverse‑engineer it. The resulting copy is then stored in a separate, GDPR‑compliant table, tagged with the shop’s identifier but devoid of personal data.

The process looks like this:

  1. Prompt Generation – Assemble brand‑only parameters in the edge function.
  2. Data Hashing – Hash any identifiers that might be needed for context.
  3. LLM Call – Send the sanitized prompt to Groq.
  4. Output Storage – Store the generated text in a table that references only the shop, not the customer.

Because the edge function runs within the same region as the Supabase database, no cross‑border data transfer occurs, and the shop can claim full control over its data.

Putting It All Together: A Sample Workflow

  1. User Interaction – A shop manager selects “Generate Promo” from the admin panel, choosing the product line and desired tone.
  2. Edge Function Invocation – The request hits the Supabase edge function, pulling the shop’s brand settings.
  3. Prompt Assembly – The function builds a concise prompt, limiting it to 35 tokens.
  4. Groq Call – The prompt is sent to the 4‑billion‑parameter LPU via Groq’s API, with max_tokens set to 30.
  5. Response Handling – The function receives a 25‑token response, validates it, and stores it in the promos table.
  6. Distribution – The admin panel queues the promo for posting to Instagram, Facebook, and the shop’s website, all within minutes.
  7. Cost Accounting – The total token count (≈65) translates to $0.000065, comfortably under the $0.01 per promo threshold.

This workflow is fully automated, requires no manual copywriting, and keeps both cost and privacy under tight control.

Conclusion

The combination of Groq’s low‑cost LPUs, prompt‑driven generation, and Supabase edge functions offers a practical path for small businesses to adopt AI‑powered marketing. By focusing on short‑form content, carefully budgeting tokens, and enforcing strict data isolation, a boutique retailer can produce fresh, brand‑aligned copy on demand, without breaking the bank or compromising privacy. The same architecture that powers the Sottocasa.ai promo generator demonstrates that this is not a theoretical exercise but a proven, production‑ready solution. For more details on how these components come together, visit the portfolio page on https://gabrielepau.com.