Add RAG Skill

Process steps in order. Do not skip ahead.

Step 1 — Add the Modules

RAG in Koog spans three concerns — each is its own module:

implementation("ai.koog:embeddings-base:1.0.0")
implementation("ai.koog:embeddings-llm:1.0.0")     // LLM-backed embeddings (OpenAI/Google/etc.)
implementation("ai.koog:rag-base:1.0.0")
implementation("ai.koog:rag-vector:1.0.0")         // vector store + similarity search

Reach past these only for backend-specific vector stores (Pinecone, Qdrant, etc.) — those usually ship as separate rag-vector-<backend> artifacts when supported.

Proceed immediately to Step 2.

Step 2 — Build the Embedder

LLM-backed embeddings via embeddings-llm use the same provider you'd use for completions:

import ai.koog.embeddings.llm.OpenAIEmbedder
import ai.koog.prompt.executor.clients.openai.OpenAIModels

val embedder = OpenAIEmbedder(
    apiKey = System.getenv("OPENAI_API_KEY"),
    model = OpenAIModels.Embedding.TextEmbedding3Small,
)

Cheaper local embedders exist for self-hosted setups — check the embeddings-* module list for what 1.0 ships beyond LLM-backed (Ollama-based embedders, in-process models).

Proceed immediately to Step 3.

Step 3 — Build the Vector Store

import ai.koog.rag.vector.InMemoryVectorStore

val store = InMemoryVectorStore(embedder = embedder)

// Index documents
documents.forEach { doc ->
    store.add(id = doc.id, text = doc.text, metadata = mapOf("source" to doc.path))
}

InMemoryVectorStore is fine for demos and small corpora. For production, swap in a persistent backend — JDBC, Pinecone, Qdrant, Weaviate — via the corresponding rag-vector-* module if shipped, or implement the VectorStore interface against your store of choice.

Proceed immediately to Step 4.

Step 4 — Wire Retrieval into the Agent

Two integration patterns:

(a) RAG as a tool — the LLM decides when to retrieve. Wrap the store in a tool:

@LLMDescription("Search the documentation corpus")
class DocsSearchTool(private val store: VectorStore) : ToolSet {
    @Tool
    @LLMDescription("Search for documents relevant to the query and return the top matches")
    suspend fun search(@LLMDescription("Search query") query: String): String {
        val results = store.search(query, topK = 5)
        return results.joinToString("\n---\n") { "${it.metadata["source"]}: ${it.text}" }
    }
}

Register in the agent's ToolRegistry. Use this pattern when retrieval is optional — some queries need it, some don't.

(b) RAG as a prompt augmenter — retrieval happens automatically before every LLM call. Use a UserPromptAugmenter (see define-prompt) that calls store.search(currentInput) and appends results to the user message. Use this pattern when retrieval is always-on for every input.

Proceed immediately to Step 5.

Step 5 — Tune Retrieval

• topK — number of documents returned. Default 5 is reasonable; raise for broader context, lower for tighter prompts
• Chunking — index full documents or split into chunks? Chunked indexing produces better recall but loses cross-chunk context. The right answer depends on document length and query shape
• Reranking — if recall is high but precision is low, add a reranker between store.search and the LLM prompt. Koog doesn't ship a built-in reranker; call a cross-encoder model or a second LLM with a scoring prompt

Finish here.