バッチ処理と非同期処理
スループットを向上させるため、埋め込み生成のバッチ処理とLLM APIの非同期呼び出しを実装します。
「バッチ処理と非同期処理」はCoddyKit上の無料LLM Apps in Production (RAG + Vector DB + Caching)レッスンです。 これはレッスン2/4です。 下記で完全なレッスンを無料で読むことができます。その後、ブラウザ内の組み込みコードエディタと24時間対応のAIチューターでハンズオン演習できます。 これはLLM Apps in Production (RAG + Vector DB + Caching)学習パスの一部であり、ウェブとCoddyKitアプリ全体で進捗が同期されます。 LLM Apps in Production (RAG + Vector DB + Caching)コースには全4レッスンが含まれています。
このレッスンの一部はまだ翻訳されておらず、英語で表示されています。
Boosting RAG Performance
Your RAG application is working, but is it fast and cost-effective? As user traffic grows, you'll need to optimize how your app interacts with LLMs and vector databases.
In this lesson, we'll explore two powerful techniques: batching and asynchronous operations. These can significantly improve your RAG system's throughput and reduce operational costs.
One Task at a Time
Imagine you have a list of documents, and you need to get embeddings for each one. In a synchronous approach, your program would process each document one by one.
- It sends a request for Document 1.
- It waits for the embedding to return.
- Then, it sends a request for Document 2.
- It waits again.
This "wait-and-process" model is simple but can be very slow, especially with many I/O operations like API calls.
Doing Things in Parallel
Asynchronous operations allow your program to send multiple requests without waiting for each one to complete before starting the next. Think of it like a restaurant manager taking multiple orders before any food is ready.
- Send request for Document 1.
- Immediately send request for Document 2.
- Immediately send request for Document 3.
- Collect results as they become available.
This approach keeps your program busy, leading to much faster overall completion times for many tasks.
How Async Works (Conceptual)
While the exact implementation varies by language, the core idea is to avoid blocking your main program thread. For example, in Java, you might use CompletableFuture to manage tasks that run in the background.
Here's a conceptual look at how you might call an LLM API asynchronously:
// Conceptual Async Call
CompletableFuture<String> futureResponse1 = llmApi.generateAsync("prompt 1");
CompletableFuture<String> futureResponse2 = llmApi.generateAsync("prompt 2");
// Do other work while responses are being generated
// Get results when ready
String response1 = futureResponse1.join();
String response2 = futureResponse2.join();This allows your application to perform other computations while waiting for I/O-bound LLM responses.
Grouping for Efficiency
Batching is the strategy of grouping several individual items or requests into a single, larger request. Many LLM providers and embedding services support batching.
- Instead of 10 separate API calls for 10 documents, make 1 API call for a "batch" of 10 documents.
- This reduces the overhead of establishing connections and processing each request individually.
- It can also be more cost-effective as some APIs charge less per token for batched requests.
Why Batch Embeddings?
Generating vector embeddings for documents is a perfect use case for batching. When you load a large dataset, you'll have many text chunks that need to be converted into vectors.
Sending these in batches to your embedding model API means:
- Fewer API calls: Reduced network latency and overhead.
- Higher throughput: Process more text in the same amount of time.
- Potential cost savings: Some APIs offer better pricing for larger batches.
Batching Embeddings Code
Let's see how to implement a simple batching mechanism for embedding generation. We'll simulate an embedding service that takes a list of texts.
This example processes a list of documents in chunks (batches) to send to an embedding service. Notice how we group documents before calling the mock service.
public class Main {
// Simulate an embedding client
static class MockEmbeddingClient {
public double[][] getEmbeddings(String[] texts) {
System.out.println("Processing batch of " + texts.length + " texts...");
// Simulate network latency
try {
Thread.sleep(100 * texts.length); // Slower for larger batches
} catch (InterruptedException e) {
Thread.currentThread().interrupt();
}
double[][] embeddings = new double[texts.length][3]; // 3D for simplicity
for (int i = 0; i < texts.length; i++) {
embeddings[i][0] = texts[i].length() * 0.1;
embeddings[i][1] = texts[i].hashCode() % 100;
embeddings[i][2] = i;
}
return embeddings;
}
}
public static void main(String[] args) {
MockEmbeddingClient client = new MockEmbeddingClient();
String[] documents = {
"The quick brown fox jumps over the lazy dog.",
"Never underestimate the power of a good book.",
"Artificial intelligence is transforming industries.",
"Retrieval Augmented Generation enhances LLM accuracy.",
"Batching improves efficiency for embedding calls.",
"Asynchronous operations prevent blocking.",
"CoddyKit makes learning fun and interactive.",
"Optimizing RAG saves costs and boosts speed."
};
int batchSize = 3;
for (int i = 0; i < documents.length; i += batchSize) {
int endIndex = Math.min(i + batchSize, documents.length);
String[] currentBatch = new String[endIndex - i];
System.arraycopy(documents, i, currentBatch, 0, endIndex - i);
double[][] batchEmbeddings = client.getEmbeddings(currentBatch);
System.out.println("Received " + batchEmbeddings.length + " embeddings for this batch.");
// In a real app, you'd store these embeddings in a vector database
}
System.out.println("All documents processed in batches.");
}
}Async + Batching = Super RAG
The real power comes from combining both techniques. You can make asynchronous calls to batched requests.
- Group your documents into batches.
- Send each batch request to the API asynchronously.
- Your program can then manage multiple concurrent batch requests, maximizing throughput.
This is crucial for ingesting massive amounts of data or handling high-volume real-time embedding lookups.
Batching & Async Pitfalls
While powerful, batching and async operations require careful handling:
- API Rate Limits: Don't send too many requests too quickly, even if they're batched. Respect API limits.
- Memory Usage: Very large batches can consume significant memory. Find an optimal batch size.
- Error Handling: If one item in a batch fails, how do you handle it? Design robust error recovery.
- Latency vs. Throughput: Async improves throughput, but might not reduce the latency of a single request.
Optimize Your Workflow
You're working on a RAG application that needs to process thousands of customer reviews to generate embeddings for a vector database. Which strategies would be most effective to improve the efficiency and speed of this data ingestion process?
Summary: Faster, Cheaper RAG
Great job! You've learned how to supercharge your RAG system's performance and cost-efficiency.
- Asynchronous operations allow concurrent processing, reducing overall wait times for I/O-bound tasks.
- Batching groups multiple small requests into larger ones, cutting down on API overhead and potentially saving costs.
- Combining both techniques provides the most powerful optimization for high-volume data processing in RAG.
These strategies are essential for building scalable and robust LLM applications in production.
よくある質問
「バッチ処理と非同期処理」レッスンは無料ですか？
はい。「バッチ処理と非同期処理」の完全なテキストはこのウェブで無料で読めます。インタラクティブに演習し（組み込みコードエディタと24時間対応のAIチューター）、LLM Apps in Production (RAG + Vector DB + Caching)コースの残りをアンロックするには、CoddyKit PROにアップグレードしてください。 LLM Apps in Production (RAG + Vector DB + Caching)コースには全4レッスンが含まれています。
「バッチ処理と非同期処理」で何を学びますか？
スループットを向上させるため、埋め込み生成のバッチ処理とLLM APIの非同期呼び出しを実装します。 ブラウザで直接実行するハンズオンコードでLLM Apps in Production (RAG + Vector DB + Caching)を演習し、24時間対応のAIチューターがレッスンを進める中での質問に答えます。
LLM Apps in Production (RAG + Vector DB + Caching)を始めるのに経験は必要ですか？
事前経験は必要ありません。CoddyKitのLLM Apps in Production (RAG + Vector DB + Caching)は初級者から上級者向けに構成されているため、ここから始めるか最初から始めて、自分のペースで進むことができます。 これはレッスン2/4です。
「バッチ処理と非同期処理」レッスンにはどのくらい時間がかかりますか？
ほとんどのCoddyKitレッスンは約5～10分かかります。各レッスンはコンパクトでインタラクティブなので、着実に進歩し、ウェブとアプリ全体で正確に前回の場所から再開できます。
このLLM Apps in Production (RAG + Vector DB + Caching)レッスンでコードを書いて実行できますか？
はい。すべてのLLM Apps in Production (RAG + Vector DB + Caching)レッスンに組み込みコードエディタが含まれているため、ブラウザでリアルコードを書いて実行し、即座のAIフィードバックを取得できます。ローカル設定は不要です。