perplexity-cost-tuning

# Perplexity Cost Tuning

## Overview
Reduce Perplexity Sonar API costs. Perplexity charges per-token (input + output) plus a per-request fee that varies by search context size. The biggest cost lever is model selection: `sonar-pro` costs 3-15x more than `sonar` per request.

## Pricing Reference

| Model | Input $/M tokens | Output $/M tokens | Request Fee |
|-------|-------------------|-------------------|-------------|
| `sonar` | $1 | $1 | $5 per 1K requests |
| `sonar-pro` | $3 | $15 | $5 per 1K requests |
| `sonar-reasoning-pro` | $3 | $15 | $5 per 1K requests |
| `sonar-deep-research` | $2 | $8 | $5 per 1K searches |

Search context size (Low/Medium/High) affects the request fee. More context = higher fee.

## Prerequisites
- Perplexity API account with usage dashboard
- Understanding of query patterns in your application
- Cache infrastructure for search results

## Instructions

### Step 1: Route Queries to the Right Model
```typescript
// 60-70% of queries can use sonar, saving 3-15x per query
function selectModel(query: string): "sonar" | "sonar-pro" {
  const simplePatterns = [
    /^what is/i, /^define/i, /^who is/i, /^when did/i,
    /current price/i, /^how many/i, /^is it true/i,
  ];
  if (simplePatterns.some((p) => p.test(query))) return "sonar";

  const complexPatterns = [
    /compare.*vs/i, /analysis of/i, /comprehensive/i,
    /pros and cons/i, /in-depth/i, /research/i,
  ];
  if (complexPatterns.some((p) => p.test(query))) return "sonar-pro";

  return "sonar"; // Default to cheapest
}
```

### Step 2: Limit Output Tokens
```bash
set -euo pipefail
# Factual queries need ~100 tokens, not 4096
# Setting max_tokens dramatically reduces output costs

# Simple fact: 100 tokens = $0.0001 output
curl -X POST https://api.perplexity.ai/chat/completions \
  -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sonar",
    "messages": [{"role": "user", "content": "Current population of Tokyo"}],
    "max_tokens": 100
  }'

# Research query: keep at 2048 only when needed
curl -X POST https://api.perplexity.ai/chat/completions \
  -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sonar-pro",
    "messages": [{"role": "user", "content": "Compare React vs Vue in 2025 for enterprise apps"}],
    "max_tokens": 2048
  }'
```

### Step 3: Cache to Eliminate Duplicate Queries
```typescript
import { LRUCache } from "lru-cache";
import { createHash } from "crypto";

const searchCache = new LRUCache<string, any>({
  max: 10000,
  ttl: 4 * 3600_000, // 4-hour default TTL
});

async function cachedQuery(query: string, model: string) {
  const key = createHash("sha256")
    .update(`${model}:${query.toLowerCase().trim()}`)
    .digest("hex");

  const cached = searchCache.get(key);
  if (cached) return cached; // $0 cost

  const result = await perplexity.chat.completions.create({
    model,
    messages: [{ role: "user", content: query }],
  });
  searchCache.set(key, result);
  return result;
}

// Track cache effectiveness
function cacheStats() {
  return {
    size: searchCache.size,
    hitRate: `${((searchCache as any).hits / ((searchCache as any).hits + (searchCache as any).misses) * 100).toFixed(1)}%`,
  };
}
```

### Step 4: Use Domain Filters to Reduce Search Cost
```bash
set -euo pipefail
# Restricting search domains = less content to process = lower request fee
curl -X POST https://api.perplexity.ai/chat/completions \
  -H "Authorization: Bearer $PERPLEXITY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sonar",
    "messages": [{"role": "user", "content": "Python 3.13 release notes"}],
    "search_domain_filter": ["python.org", "docs.python.org"],
    "max_tokens": 500
  }'
```

### Step 5: Track and Budget
```typescript
class CostTracker {
  private costs: Array<{ model: string; tokens: number; timestamp: Date }> = [];

  record(model: string, usage: { total_tokens: number }) {
    this.costs.push({
      model,
      tokens: usage.total_tokens,
      timestamp: new Date(),
    });
  }

  dailySummary() {
    const today = this.costs.filter(
      (c) => c.timestamp.toDateString() === new Date().toDateString()
    );
    const sonarTokens = today.filter((c) => c.model === "sonar").reduce((s, c) => s + c.tokens, 0);
    const proTokens = today.filter((c) => c.model === "sonar-pro").reduce((s, c) => s + c.tokens, 0);

    return {
      queries: today.length,
      estimatedCost: (sonarTokens * 0.000001) + (proTokens * 0.000009), // rough estimate
      sonarQueries: today.filter((c) => c.model === "sonar").length,
      proQueries: today.filter((c) => c.model === "sonar-pro").length,
    };
  }
}
```

## Cost Optimization Checklist
- [ ] Default model is `sonar` (not `sonar-pro`)
- [ ] `max_tokens` set on every request
- [ ] Caching enabled for repeated queries
- [ ] Model routing by query complexity
- [ ] Domain filter used where applicable
- [ ] Monthly budget cap set on API key
- [ ] Cost tracking in production monitoring

## Error Handling
| Issue | Cause | Solution |
|-------|-------|----------|
| High cost per query | Using sonar-pro for everything | Route simple queries to sonar |
| Low cache hit rate | Queries too unique | Normalize queries before hashing |
| Budget exhausted early | No spending caps | Set monthly budget on API key |
| Unexpectedly high bill | No max_tokens limits | Set max_tokens on all requests |

## Output
- Model routing saving 60-70% on simple queries
- Token limiting reducing output costs
- Caching eliminating duplicate query costs
- Cost tracking for budget monitoring

## Resources
- [Perplexity Pricing](https://docs.perplexity.ai/docs/getting-started/pricing)
- [Model Cards](https://docs.perplexity.ai/getting-started/models)

## Next Steps
For architecture patterns, see `perplexity-reference-architecture`.