Document Upload & Indexing for RAG in Mobile Apps

The app crashes when loading PDFs over 50 MB — a typical complaint during QA. 90% of the time, loading directly into memory causes OutOfMemoryError on Android and poor performance on iOS. We solved this with streaming data and an async pipeline. The user attaches a PDF from Files.app or gallery, taps "Upload", and within seconds can ask questions about the document. Behind these seconds is a complex pipeline: file upload to server, parsing, chunking, embedding generation, and writing to a vector DB. Each step is a potential source of problems: memory leaks on mobile, connection drops, incorrect chunking, data loss. We have implemented such a pipeline end-to-end for 5 projects, including apps with millions of documents. Our experience shows that without proper architecture, the pipeline can collapse under load. This article details the technical solutions for each step — from selecting the upload API to error handling on the server.

How to Avoid Memory Issues During Document Upload?

Android: Choosing the Right API

For Android 13+ we use ActivityResultContracts.GetContent() with MIME type application/pdf or */*. We get a Uri of type content://. To read data without OutOfMemoryError for large files, we don't load everything into ByteArray — we pass InputStream directly via a custom RequestBody in OkHttp. Example:

val uri: Uri = // from ActivityResult
val requestBody = object : RequestBody() {
    override fun contentType() = "application/octet-stream".toMediaType()
    override fun contentLength() = -1 // unknown
    override fun writeTo(sink: BufferedSink) {
        contentResolver.openInputStream(uri)?.use { input ->
            sink.writeAll(input.source())
        }
    }
}

For files over 50 MB — chunked upload via MultipartBody calling writeTo in parts. This avoids RAM overload and improves app stability.

iOS: UIDocumentPicker and Streaming Upload

On iOS we use UIDocumentPickerViewController with UTType.pdf etc. We get a URL of type file://. Reading via Data(contentsOf:) for files >20 MB is dangerous — better to use URLSession.uploadTask(with:fromFile:), which reads the file directly from the filesystem without loading into memory.

let request = URLRequest(url: uploadEndpoint)
let (_, response) = try await URLSession.shared.upload(for: request, fromFile: fileURL)

More about URLSession — Apple's standard documentation.

Upload Progress

A progress bar is mandatory: on Android — via CountingOutputStream in OkHttp, on iOS — via URLSession's uploadProgress property. The user should see how much is left. Without progress, large files feel like they are hanging. Typical upload time for a 20-30 MB file is about 5 seconds on a good connection.

Step 2: Server Pipeline — File to Vectors

The process includes: 1) file upload, 2) text extraction, 3) chunking, 4) embedding generation, 5) vector DB insertion. After the file is uploaded to the server, we start an async pipeline via Celery (or similar queue). The client immediately gets {"job_id": "abc", "status": "processing"}, and status is tracked via polling or WebSocket.

@celery.task
def process_document(file_path: str, user_id: str, content_type: str):
    text = extract_text(file_path, content_type)
    chunks = split_into_chunks(text, chunk_size=500, overlap=50)
    embeddings = create_embeddings_batch(chunks)
    upsert_to_vector_store(chunks, embeddings, user_id)
    update_document_status(file_path, "completed")

Document Parsing — table:

PyMuPDF is our choice for PDF: correctly handles Cyrillic, preserves font information (useful for heading detection). Source: PyMuPDF.

Step 3: Chunking Strategy — Fixed Size vs Semantic

Fixed chunking is faster and more predictable, but can break semantic blocks. Semantic is slower but improves RAG answer quality by 15–20% in our measurements. Chunk size depends on the embedding model: for text-embedding-3-small optimal is 512 tokens. If the document contains tables or code, increase overlap to 100 tokens. Using semantic chunking with transformer-based models can boost accuracy by 20-30%.

Step 4: Tracking Progress and Indexing Status

While the document is processing, the client should see the current state. Two approaches:

• Polling — every 2–3 seconds query /api/documents/{job_id}/status. Simple, works everywhere, but generates 20-30 extra requests per minute.
• WebSocket / SSE — client subscribes to events for job_id. Backend sends updates: {"step": "chunking", "progress": 0.3} → {"step": "embedding", "progress": 0.7} → {"step": "completed"}. Best UX, but harder to implement with background work.

We use SSE as a balance between complexity and user experience. After indexing completes — push notification via APNs/FCM. We reduced indexing time for one document by 40% through parallel chunk processing.

Step 5: User Document Management

The user can delete documents. When deleting:

1. Delete the file from storage.
2. Delete all chunks from vector DB (by user_id + document_id).
3. Update the record in the relational DB.

In Pinecone: index.delete(filter={"document_id": "xyz"}, namespace=user_id). In pgvector: DELETE FROM documents WHERE document_id = $1 AND user_id = $2.

Documents are stored with metadata: name, size, upload date, chunk count, status. This allows recovery in case of failure.

What's Included in Our RAG Pipeline Implementation

We provide:

• Pipeline architecture design for your scenarios
• Upload implementation with progress on iOS/Android
• Server side: parsing, chunking, embeddings, vector storage
• Integration with Firebase Cloud Messaging for notifications
• API documentation and deployment instructions
• Team training on the system
• Stability guarantee: unit tests and load testing

Our team has over 5 years of experience in mobile development and has successfully implemented RAG pipelines for 5+ projects. Get a consultation on pipeline architecture.

Timeline and Cost

MVP with PDF and TXT support, basic chunking, and pgvector — 3–4 weeks (from $5,000). Full pipeline with OCR, multiple formats, async queue, and WebSocket status — 6–8 weeks (from $12,000). Cost is calculated individually after requirements analysis.

Contact us for a project assessment. Order a turnkey implementation.

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.