Implement Token-Aware Prompts to Avoid Wasted Context

LLM pricing is based on tokens (input + output). Sending a 10,000-token document to GPT-4o when you only need information from the first paragraph wastes money. Pre-process your inputs to reduce token count before sending them to the LLM.

Real-world example: A workflow summarizes customer call transcripts. Some transcripts are 30 pages long, but the relevant content is usually in the first 5 pages and the last 2 pages (intro and conclusion). Trim the middle to save tokens.

// Code node: "Optimize LLM Input"
const transcript = $input.first().json.transcript;
const maxChars = 8000; // Roughly 2000 tokens for English text

// Strategy 1: Simple truncation with notice
if (transcript.length > maxChars) {
  const truncated = transcript.substring(0, maxChars);
  return [{
    json: {
      optimizedText: truncated + '\n\n[Transcript truncated for length]',
      originalLength: transcript.length,
      optimizedLength: truncated.length,
      estimatedTokensSaved: Math.floor((transcript.length - truncated.length) / 4)
    }
  }];
}

return [{
  json: {
    optimizedText: transcript,
    originalLength: transcript.length,
    optimizedLength: transcript.length,
    estimatedTokensSaved: 0
  }
}];
```text
A smarter approach -- keep the beginning and end, summarize the middle:

```javascript
// Code node: "Smart Transcript Trimming"
const transcript = $input.first().json.transcript;
const maxChars = 12000;

if (transcript.length <= maxChars) {
  return [{ json: { optimizedText: transcript, strategy: 'none' } }];
}

// Keep first 4000 chars (intro, context)
const head = transcript.substring(0, 4000);

// Keep last 3000 chars (conclusion, action items)
const tail = transcript.substring(transcript.length - 3000);

// Middle section -- will be sent to a cheap model for compression
const middle = transcript.substring(4000, transcript.length - 3000);

return [{
  json: {
    head,
    middle,
    tail,
    strategy: 'head-tail-with-middle-summary',
    estimatedTokensSaved: Math.floor(middle.length / 4)
  }
}];
```text
Then use a two-stage LLM pipeline:

```yaml
Stage 1 (cheap model - gpt-4o-mini):
  Input: The middle section
  Prompt: "Summarize the key points from this section in 3-5 bullet points."
  Cost: ~$0.001

Stage 2 (full model - gpt-4o):
  Input: head + middle_summary + tail
  Prompt: "Provide a comprehensive summary of this customer call."
  Cost: ~$0.01 (instead of $0.05 with full transcript)
```text
```yaml
Savings per transcript:
  Full 30-page transcript to gpt-4o: ~$0.05
  Optimized (head + summary + tail): ~$0.011
  Savings per call: ~78%
  At 100 calls/day: $150/month saved
```text
Always profile your actual token usage. Add a logging step that records input tokens, output tokens, and cost per LLM call so you can identify the most expensive workflows.

**Related:** [Use Structured Output (JSON Mode) for Parseable Responses](../ai-and-llm-integration/01-use-structured-output-json-mode-for-parseable-responses.md) | [Configure Payload Size and Binary Data Mode for Large Files](../performance-and-large-files/01-configure-payload-size-and-binary-data-mode-for-large-files.md)