Prose Balance Review with Nebius Llama-3.3-70B — Structural Writing Analysis at $0.10/1M

#flutter #supabase #ai #buildinpublic

Prose Balance Review with Nebius Llama-3.3-70B

What is "Balance Review"?

Balance review goes beyond spell-check. It detects structural imbalances in writing:

One paragraph significantly longer than others
Bullet point granularity inconsistent across a list
Weak conclusion or overlong introduction

This requires moderate reasoning ability — more than a tag suggester, less than a full editorial rewrite. Nebius's llama-3.3-70b at $0.10/1M tokens is the right fit.

Where It Fits in the AI Routing Table

Task	Model	Cost
Tag suggestions	Groq llama-3.3-70b	Free tier
Bulk summarization	DeepInfra llama-3.1-70b	$0.07/1M
Prose balance review	Nebius llama-3.3-70b	$0.10/1M
Document-level improvements	Claude Haiku	$0.25/1M
Design decisions	Claude Sonnet	$3.00/1M

Nebius sits between DeepInfra (cost-optimized) and Claude Haiku (quality-optimized).

Supabase Edge Function

// ai-hub/index.ts (action: "notes.balance_review")
case "notes.balance_review": {
  const { content } = body;

  const response = await fetch(
    "https://api.studio.nebius.ai/v1/chat/completions",
    {
      method: "POST",
      headers: {
        "Authorization": `Bearer ${Deno.env.get("NEBIUS_API_KEY")}`,
        "Content-Type": "application/json",
      },
      body: JSON.stringify({
        model: "meta-llama/Llama-3.3-70B-Instruct",
        messages: [
          {
            role: "system",
            content: `Analyze the writing balance and suggest up to 3 improvements.
Return JSON: {"issues": [{"type": "paragraph_balance|list_granularity|conclusion_strength", "description": "...", "suggestion": "..."}]}`,
          },
          { role: "user", content: content.slice(0, 2000) },
        ],
        max_tokens: 300,
        temperature: 0.3,
        response_format: { type: "json_object" },
      }),
    }
  );

  const data = await response.json();
  const result = JSON.parse(data.choices[0].message.content);

  return new Response(JSON.stringify(result), {
    headers: { "Content-Type": "application/json" },
  });
}

response_format: { type: "json_object" } enforces structured output — no parse errors.

Flutter Side

// note_editor_page.dart
Future<void> _requestBalanceReview() async {
  setState(() => _reviewLoading = true);

  try {
    final response = await Supabase.instance.client.functions.invoke(
      'ai-hub',
      body: {
        'action': 'notes.balance_review',
        'content': _controller.text,
      },
    );

    final issues = (response.data['issues'] as List? ?? [])
        .cast<Map<String, dynamic>>();

    if (mounted) {
      setState(() {
        _reviewIssues = issues;
        _reviewLoading = false;
      });
    }
  } catch (e) {
    setState(() => _reviewLoading = false);
  }
}

A "Balance Check" button below the editor triggers this. Results render as issue cards.

Nebius Quick Reference

Property	Value
Endpoint	`api.studio.nebius.ai/v1/` (OpenAI-compatible)
Models	Llama-3.3-70B, DeepSeek-V3, others
Input cost	$0.10/1M tokens
JSON mode	✅ supported
Japanese quality	Solid at 70B scale

The Routing Pattern

The consistent theme across Groq, DeepInfra, and Nebius: all use OpenAI-compatible APIs. Your routing layer is just a URL swap:

const PROVIDER_URLS = {
  groq: "https://api.groq.com/openai/v1",
  deepinfra: "https://api.deepinfra.com/v1/openai",
  nebius: "https://api.studio.nebius.ai/v1",
};