How Many Log Lines Should You Send to Gemini? The Context Window Problem.

#ai #gemini #programming #android

All tests run on an 8-year-old MacBook Air.

When you send logcat lines to Gemini for diagnosis, how many lines should you include?

Too few: the model lacks context, gives a generic answer.
Too many: you hit token limits, slow down the response, and bury the relevant signal in noise.

Here's what I found after testing different window sizes in HiyokoLogcat.

The naive approach: send everything

First attempt: grab the last 500 lines and send them all.

Problems:

Gemini 1.5 Flash has a large context window, but 500 verbose logcat lines is a lot of tokens
Response latency increases noticeably
The model sometimes focuses on unrelated noise rather than the actual error

What actually works: 50 lines before, 50 after

After testing, ±50 lines around the target error hits the sweet spot:

pub fn get_diagnosis_context(
    buffer: &LogRingBuffer,
    error_idx: usize,
) -> String {
    let window = 50;
    let lines = buffer.context_around(error_idx, window);

    lines
        .iter()
        .map(|l| format!("[{}] {}/{}: {}",
            l.timestamp, l.level, l.tag, l.message))
        .collect::>()
        .join("\n")
}

50 lines before captures the sequence of events leading to the crash.
50 lines after captures any follow-on errors and recovery attempts.

100 lines total. Typically 3,000–6,000 tokens. Fast response, relevant diagnosis.

When to go wider

Some crashes only make sense with more context — memory leaks that build up over minutes, threading issues with long setup chains.

For these, I added a "deep diagnosis" mode that sends ±200 lines:

pub enum DiagnosisMode {
    Quick,  // ±50 lines
    Deep,   // ±200 lines
}

Quick is the default. Deep is one extra click. Most users never need it.

The format matters too

Raw logcat output has a lot of repetitive structure. Reformatting before sending reduces token count without losing information:

// Raw (verbose)
04-20 10:23:45.123  1234  5678 E AndroidRuntime: FATAL EXCEPTION: main

// Reformatted (compact)
[10:23:45] E/AndroidRuntime: FATAL EXCEPTION: main

Same information, fewer tokens. Over 100 lines this adds up.

The result

±50 lines, reformatted, PII-masked → Gemini gives a specific, actionable diagnosis in 2–3 seconds on the free tier.

That's the configuration I shipped with.

HiyokoLogcat is free and open source → github.com/hiyoyok/HiyokoLogcat
X → @hiyoyok

DEV Community