กระบวนการสร้างแบบเสริมด้วยการดึงข้อมูล

Retrieval-Augmented Generation (RAG) grounds an LLM's answers in your own documents instead of relying solely on what the model memorized during training.

• Fresh & private data — answer questions about internal docs the model never saw.
• Less hallucination — the model cites retrieved context rather than inventing facts.
• Cheaper than fine-tuning — you update a vector store, not model weights.

A Spring AI RAG pipeline has two phases: an ingestion phase (read → split → embed → store) and a query phase (embed question → retrieve → augment prompt → generate).

Spring AI gives you composable building blocks for both phases. The core types you will assemble are:

• DocumentReader — loads raw sources (PDF, Markdown, JSON, web pages).
• DocumentTransformer — splits documents into chunks (e.g. TokenTextSplitter).
• EmbeddingModel — turns text into vectors.
• VectorStore — stores and similarity-searches those vectors.
• ChatClient with a RAG advisor — wires retrieval into the prompt automatically.

The first three feed ingestion; the last two power querying.

During ingestion you read source files and split them into chunks small enough to fit the model's context window while staying semantically coherent. TokenTextSplitter chunks by token count with overlap so meaning isn't cut mid-sentence.

Below, a Markdown file is read and split into ~800-token chunks before storage.

An embedding is a dense vector that captures the meaning of text. Two passages about the same topic land close together in vector space, which is what makes similarity search work.

Spring AI auto-configures an EmbeddingModel based on your starter (OpenAI, Azure, Ollama, etc.). You rarely call it directly — the VectorStore uses it internally — but you can:

The VectorStore persists embeddings and runs similarity searches. Spring AI ships adapters for PgVector, Redis, Chroma, Qdrant, Milvus, and a simple in-memory store.

With the spring-ai-starter-vector-store-pgvector dependency, a bean is auto-configured from properties — no manual wiring needed:

At query time you turn the user's question into a vector and ask the store for the nearest chunks. A SearchRequest controls how many results (topK) and a minimum similarityThreshold to filter out weak matches.

Before reaching for advisors, it helps to see the mechanic. RAG simply stuffs the retrieved text into the prompt and instructs the model to answer only from it. This is the "augment" step:

Spring AI packages that manual pattern as the QuestionAnswerAdvisor. You attach it to a ChatClient and it transparently runs the similarity search and injects context for every call.

This is the idiomatic, declarative way to do RAG in Spring Boot 4:

Defaults are convenient, but real apps tune retrieval per request — a focused FAQ may need topK=2, a broad research query topK=8. Pass an advisor at call time and override its SearchRequest, plus metadata filters to scope to a tenant or document set:

For advanced pipelines, Spring AI 1.0 offers the Modular RAG API via RetrievalAugmentationAdvisor. It exposes each stage as a swappable component:

• QueryTransformer — rewrite, compress, or translate the query.
• DocumentRetriever — the source of chunks (e.g. VectorStoreDocumentRetriever).
• QueryAugmenter — control how context is merged and how empty-context is handled.

A subtle RAG failure: when retrieval finds nothing relevant, a naive prompt lets the model fall back to its training data and hallucinate. The ContextualQueryAugmenter lets you decide that policy explicitly.

Set allowEmptyContext(false) to force a graceful "I don't have information on that" instead of a confident guess:

You build a RAG assistant. When a user asks something outside your knowledge base, similarity search returns no chunks above the threshold, yet the model still answers confidently with made-up facts. Which change best fixes this?

You assembled a complete RAG pipeline in Spring AI:

• Ingestion — DocumentReader → TokenTextSplitter → EmbeddingModel → VectorStore.add().
• Query — embed the question, similaritySearch with topK and similarityThreshold, then augment the prompt.
• Declarative RAG — QuestionAnswerAdvisor wires retrieval into a ChatClient automatically; tune it per request with custom SearchRequest and metadata filters.
• Modular RAG — RetrievalAugmentationAdvisor with query transformers, retrievers, and augmenters for advanced control.
• Safety — ContextualQueryAugmenter.allowEmptyContext(false) prevents hallucination when nothing relevant is retrieved.

Grounding model output in retrieved context is the single most effective way to make LLM features trustworthy.