AI Summarization of Long Texts in Mobile Apps

AI Summarization of Long Texts: Fitting Everything into the Context Window

Summarizing a long text immediately hits the context window limit. GPT-4o accepts 128K tokens (roughly 100K words), Claude 3 — 200K. At first glance, that's enough, but legal contracts of 200 pages, technical reports, or books often exceed the limit. Even if the text fits, a long context increases cost and response latency. For example, a typical supply contract may contain 150 pages of fine print — a direct request would cost about $0.03 to process, but the result may be incomplete due to the "lost in the middle" effect. Research on Lost in the Middle shows that models recall less information from the middle of a document. Therefore, for large documents, decomposition strategies are needed.

We use three main strategies: direct summarization, Map-Reduce, and Refine. Each suits a different size and quality requirement. Below are the details.

Strategies for Different Text Lengths

Direct summarization works for texts up to 50–80K tokens. We send the entire text in one request and ask for a summary. Simple, cheap to implement. The limitation is token cost and latency (the model processes a large context more slowly).

Map-Reduce is for texts that don't fit into the context. We split into chunks → summarize each chunk → summarize the summaries. Map-Reduce is 3 times faster than Refine for large documents because it processes chunks in parallel.

async def map_reduce_summarize(text: str, chunk_size: int = 4000) -> str:
    chunks = split_text(text, chunk_size)
    chunk_summaries = await asyncio.gather(*[
        summarize_chunk(chunk) for chunk in chunks
    ])
    combined = "\n\n".join(chunk_summaries)
    if count_tokens(combined) > chunk_size:
        return await map_reduce_summarize(combined, chunk_size)
    return await summarize_final(combined)

asyncio.gather — parallel API requests for all chunks simultaneously. For 10 chunks, the time is nearly the same as for one.

Refine — summarize the first chunk, then refine the summary with each subsequent chunk. The final summary is enriched sequentially. Quality is higher than Map-Reduce for connected narrative texts, but slower — requests are sequential.

How to Manage Tokens and Avoid Errors?

The main mistake is not counting tokens before sending. tiktoken (Python) or gpt-tokenizer (JS) give accurate counts:

import tiktoken

enc = tiktoken.encoding_for_model("gpt-4o")
token_count = len(enc.encode(text))

if token_count < 100_000:
    return await direct_summarize(text)
elif token_count < 500_000:
    return await map_reduce_summarize(text, chunk_size=8000)
else:
    return await map_reduce_summarize(text, chunk_size=4000)

Different summary types require different prompts:

• Executive summary (for executives): 3–5 sentences, only key decisions and figures
• Detailed retelling: structured list with subtitles
• Key points list: bullets without narrative
• Answer to a question: "what is this document and what needs to be done?"

On mobile (using Swift or Kotlin), we offer the user to choose the summary type before launch.

How to Avoid Duplication in Summarization?

With Map-Reduce, the final summary may repeat similar points from different chunks. Duplication is eliminated by explicitly stating in the prompt: "Combine similar points, do not repeat the same idea twice." For legal and financial documents, we use structured output in JSON format with fixed fields (parties, obligations, deadlines, key_figures). This is more reliable than free text.

How to Configure Summarization for Your Document

1. Determine the model's maximum context size (e.g., 128K for GPT-4o).
2. Split the text into chunks of 4–8K tokens, respecting paragraph boundaries.
3. Choose a strategy: direct for short texts, Map-Reduce for medium, Refine for connected narratives.
4. Configure prompts for each summary type (executive, detailed, etc.).
5. Implement caching by document hash and progress updates via SSE.

Summarization Progress on Mobile

Summarizing a 100-page document takes 15–60 seconds. Without a progress indicator, the UX suffers. The server sends events via SSE:

event: progress
data: {"step": "chunking", "total_chunks": 12, "completed": 0}

event: progress
data: {"step": "summarizing", "total_chunks": 12, "completed": 4}

event: result
data: {"summary": "...", "word_count": 450}

On the mobile client, a progress bar with a step description and animated text "Processing pages 1–25...".

Caching

Summarizing a document costs money. Cache the result by content hash + summary type. Redis with a TTL of 7–30 days is standard. If the document changes, invalidate the cache by document_id.

What Our Work Includes

• Task analysis and strategy selection (direct, Map-Reduce, Refine)
• Server pipeline development with model integration
• Caching and streaming setup
• Mobile UI with summary type selection and progress bar
• Testing on real customer documents
• Detailed documentation, system access, team training, and 2 weeks of post-launch support

Phases and Timelines

With over 5 years of experience and 30+ AI projects delivered, we guarantee reliable summarization. This solution saves companies up to $10,000 annually in manual summarization costs. Our certified AI expertise and secure infrastructure ensure your data remains confidential. We'll assess your project in one day — contact us for a consultation. Get a demo in 2–3 days.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.