Corrective RAG (CRAG) with LangGraph4j¶
This how-to ports the historical LangGraph Corrective RAG tutorial to Java with LangGraph4j and LangChain4j.
Corrective Retrieval Augmented Generation evaluates retrieved knowledge before generation. The paper also describes confidence-based actions and knowledge refinement. Like the original tutorial, this introductory implementation intentionally uses a smaller workflow: grade every retrieved document, keep relevant local documents, and use a rewritten query plus web search when retrieval is empty or any document is irrelevant.
START -> retrieve -> grade_documents -> generate -> END
|
+-> transform_query -> web_search -> generate -> END
Setup¶
Install LangGraph4j in your local Maven repository, then configure an OpenAI-compatible service that provides both chat and embeddings. OpenAI works with the default model names. For Qwen, set OPENAI_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1, OPENAI_CHAT_MODEL=qwen-plus, and OPENAI_EMBEDDING_MODEL=text-embedding-v4. If Model Studio displays a workspace-specific endpoint for your region, use that value instead. Set OPENAI_API_KEY and TAVILY_API_KEY before running the notebook.
In [ ]:
Copied!
var userHomeDir = System.getProperty("user.home");
var localRepoUrl = "file://" + userHomeDir + "/.m2/repository/";
var langgraph4jVersion = "1.9.beta4";
var langchain4jVersion = "1.19.0";
var langchain4jBetaVersion = "1.19.0-beta29";
var userHomeDir = System.getProperty("user.home");
var localRepoUrl = "file://" + userHomeDir + "/.m2/repository/";
var langgraph4jVersion = "1.9.beta4";
var langchain4jVersion = "1.19.0";
var langchain4jBetaVersion = "1.19.0-beta29";
In [2]:
Copied!
%dependency /add-repo local \{localRepoUrl} release|never snapshot|always
%dependency /add org.slf4j:slf4j-jdk14:2.0.9
%dependency /add org.bsc.langgraph4j:langgraph4j-langchain4j:\{langgraph4jVersion}
%dependency /add dev.langchain4j:langchain4j:\{langchain4jVersion}
%dependency /add dev.langchain4j:langchain4j-open-ai:\{langchain4jVersion}
%dependency /add dev.langchain4j:langchain4j-web-search-engine-tavily:\{langchain4jBetaVersion}
%dependency /resolve
%dependency /add-repo local \{localRepoUrl} release|never snapshot|always
%dependency /add org.slf4j:slf4j-jdk14:2.0.9
%dependency /add org.bsc.langgraph4j:langgraph4j-langchain4j:\{langgraph4jVersion}
%dependency /add dev.langchain4j:langchain4j:\{langchain4jVersion}
%dependency /add dev.langchain4j:langchain4j-open-ai:\{langchain4jVersion}
%dependency /add dev.langchain4j:langchain4j-web-search-engine-tavily:\{langchain4jBetaVersion}
%dependency /resolve
Repository local url: file:///Users/bsorrentino/.m2/repository/ added. Adding dependency org.slf4j:slf4j-jdk14:2.0.9 Adding dependency org.bsc.langgraph4j:langgraph4j-langchain4j:1.8.26 Adding dependency dev.langchain4j:langchain4j:1.19.0 Adding dependency dev.langchain4j:langchain4j-open-ai:1.19.0 Adding dependency dev.langchain4j:langchain4j-web-search-engine-tavily:1.19.0-beta29 Solving dependencies Resolved artifacts count: 20 Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/slf4j/slf4j-jdk14/2.0.9/slf4j-jdk14-2.0.9.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/slf4j/slf4j-api/2.0.9/slf4j-api-2.0.9.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/bsc/langgraph4j/langgraph4j-langchain4j/1.8.26/langgraph4j-langchain4j-1.8.26.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/dev/langchain4j/langchain4j-skills/1.12.1-beta21/langchain4j-skills-1.12.1-beta21.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/commonmark/commonmark/0.26.0/commonmark-0.26.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/commonmark/commonmark-ext-yaml-front-matter/0.26.0/commonmark-ext-yaml-front-matter-0.26.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/bsc/langgraph4j/langgraph4j-core/1.8.26/langgraph4j-core-1.8.26.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/bsc/async/async-generator/4.3.1/async-generator-4.3.1.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/dev/langchain4j/langchain4j/1.19.0/langchain4j-1.19.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/dev/langchain4j/langchain4j-core/1.19.0/langchain4j-core-1.19.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/jspecify/jspecify/1.0.0/jspecify-1.0.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/com/fasterxml/jackson/core/jackson-annotations/2.22/jackson-annotations-2.22.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/com/fasterxml/jackson/core/jackson-core/2.22.1/jackson-core-2.22.1.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/com/fasterxml/jackson/core/jackson-databind/2.22.1/jackson-databind-2.22.1.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/org/apache/opennlp/opennlp-tools/2.5.11/opennlp-tools-2.5.11.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/dev/langchain4j/langchain4j-open-ai/1.19.0/langchain4j-open-ai-1.19.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/dev/langchain4j/langchain4j-http-client/1.19.0/langchain4j-http-client-1.19.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/dev/langchain4j/langchain4j-http-client-jdk/1.19.0/langchain4j-http-client-jdk-1.19.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/com/knuddels/jtokkit/1.1.0/jtokkit-1.1.0.jar Add to classpath: /Users/bsorrentino/Library/Jupyter/kernels/rapaio-jupyter-kernel/mima_cache/dev/langchain4j/langchain4j-web-search-engine-tavily/1.19.0-beta29/langchain4j-web-search-engine-tavily-1.19.0-beta29.jar
In [3]:
Copied!
var openAiKey = System.getenv("OPENAI_API_KEY");
var openAiBaseUrl = System.getenv("OPENAI_BASE_URL");
var chatModelName = System.getenv().getOrDefault("OPENAI_CHAT_MODEL", "gpt-4o-mini");
var embeddingModelName = System.getenv().getOrDefault("OPENAI_EMBEDDING_MODEL", "text-embedding-3-small");
var tavilyKey = System.getenv("TAVILY_API_KEY");
if (openAiKey == null || openAiKey.isBlank()) {
throw new IllegalStateException("Set OPENAI_API_KEY before running this notebook");
}
if (tavilyKey == null || tavilyKey.isBlank()) {
throw new IllegalStateException("Set TAVILY_API_KEY before running this notebook");
}
var openAiKey = System.getenv("OPENAI_API_KEY");
var openAiBaseUrl = System.getenv("OPENAI_BASE_URL");
var chatModelName = System.getenv().getOrDefault("OPENAI_CHAT_MODEL", "gpt-4o-mini");
var embeddingModelName = System.getenv().getOrDefault("OPENAI_EMBEDDING_MODEL", "text-embedding-3-small");
var tavilyKey = System.getenv("TAVILY_API_KEY");
if (openAiKey == null || openAiKey.isBlank()) {
throw new IllegalStateException("Set OPENAI_API_KEY before running this notebook");
}
if (tavilyKey == null || tavilyKey.isBlank()) {
throw new IllegalStateException("Set TAVILY_API_KEY before running this notebook");
}
Build the local index¶
Load the same three blog posts used by the original tutorial and index them in an in-memory embedding store.
In [4]:
Copied!
import dev.langchain4j.data.document.Document;
import dev.langchain4j.data.document.loader.UrlDocumentLoader;
import dev.langchain4j.data.document.parser.TextDocumentParser;
import dev.langchain4j.data.document.splitter.DocumentSplitters;
import dev.langchain4j.data.segment.TextSegment;
import dev.langchain4j.model.openai.OpenAiEmbeddingModel;
import dev.langchain4j.rag.content.retriever.EmbeddingStoreContentRetriever;
import dev.langchain4j.store.embedding.EmbeddingStoreIngestor;
import dev.langchain4j.store.embedding.inmemory.InMemoryEmbeddingStore;
import java.util.List;
var urls = List.of(
"https://lilianweng.github.io/posts/2023-06-23-agent/",
"https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/",
"https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm/"
);
var sourceDocuments = urls.stream()
.map(url -> UrlDocumentLoader.load(url, new TextDocumentParser()))
.toList();
var embeddingModelBuilder = OpenAiEmbeddingModel.builder()
.apiKey(openAiKey)
.modelName(embeddingModelName)
// Qwen text-embedding-v4 accepts at most 10 inputs per request.
.maxSegmentsPerBatch(10);
// A custom base URL lets the same tutorial run against Qwen or another compatible service.
if (openAiBaseUrl != null && !openAiBaseUrl.isBlank()) {
embeddingModelBuilder.baseUrl(openAiBaseUrl);
}
var embeddingModel = embeddingModelBuilder.build();
var embeddingStore = new InMemoryEmbeddingStore<TextSegment>();
EmbeddingStoreIngestor.builder()
.documentSplitter(DocumentSplitters.recursive(250, 0))
.embeddingModel(embeddingModel)
.embeddingStore(embeddingStore)
.build()
.ingest(sourceDocuments);
var retriever = EmbeddingStoreContentRetriever.builder()
.embeddingStore(embeddingStore)
.embeddingModel(embeddingModel)
.maxResults(3)
.build();
import dev.langchain4j.data.document.Document;
import dev.langchain4j.data.document.loader.UrlDocumentLoader;
import dev.langchain4j.data.document.parser.TextDocumentParser;
import dev.langchain4j.data.document.splitter.DocumentSplitters;
import dev.langchain4j.data.segment.TextSegment;
import dev.langchain4j.model.openai.OpenAiEmbeddingModel;
import dev.langchain4j.rag.content.retriever.EmbeddingStoreContentRetriever;
import dev.langchain4j.store.embedding.EmbeddingStoreIngestor;
import dev.langchain4j.store.embedding.inmemory.InMemoryEmbeddingStore;
import java.util.List;
var urls = List.of(
"https://lilianweng.github.io/posts/2023-06-23-agent/",
"https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/",
"https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm/"
);
var sourceDocuments = urls.stream()
.map(url -> UrlDocumentLoader.load(url, new TextDocumentParser()))
.toList();
var embeddingModelBuilder = OpenAiEmbeddingModel.builder()
.apiKey(openAiKey)
.modelName(embeddingModelName)
// Qwen text-embedding-v4 accepts at most 10 inputs per request.
.maxSegmentsPerBatch(10);
// A custom base URL lets the same tutorial run against Qwen or another compatible service.
if (openAiBaseUrl != null && !openAiBaseUrl.isBlank()) {
embeddingModelBuilder.baseUrl(openAiBaseUrl);
}
var embeddingModel = embeddingModelBuilder.build();
var embeddingStore = new InMemoryEmbeddingStore<TextSegment>();
EmbeddingStoreIngestor.builder()
.documentSplitter(DocumentSplitters.recursive(250, 0))
.embeddingModel(embeddingModel)
.embeddingStore(embeddingStore)
.build()
.ingest(sourceDocuments);
var retriever = EmbeddingStoreContentRetriever.builder()
.embeddingStore(embeddingStore)
.embeddingModel(embeddingModel)
.maxResults(3)
.build();
Define the graph state¶
A small serializable record represents both local and web results. The grading node replaces the retrieved list with its filtered list, while the web-search node explicitly merges new results.
In [5]:
Copied!
import org.bsc.langgraph4j.state.AgentState;
import org.bsc.langgraph4j.state.Channel;
import org.bsc.langgraph4j.state.Channels;
import java.io.Serializable;
import java.util.ArrayList;
import java.util.Map;
import java.util.Optional;
record RagDocument(String content, String source) implements Serializable {}
class CorrectiveRagState extends AgentState {
static final String QUESTION = "question";
static final String DOCUMENTS = "documents";
static final String WEB_SEARCH = "web_search";
static final String GENERATION = "generation";
static final Map<String, Channel<?>> SCHEMA = Map.of(
QUESTION, Channels.base(() -> ""),
DOCUMENTS, Channels.base(ArrayList::new),
WEB_SEARCH, Channels.base(() -> false),
GENERATION, Channels.base(() -> "")
);
CorrectiveRagState(Map<String, Object> initData) { super(initData); }
String question() { return this.<String>value(QUESTION).orElse(""); }
@SuppressWarnings("unchecked")
List<RagDocument> documents() {
return this.<List<RagDocument>>value(DOCUMENTS).orElse(List.of());
}
boolean webSearch() { return this.<Boolean>value(WEB_SEARCH).orElse(false); }
Optional<String> generation() { return value(GENERATION); }
}
import org.bsc.langgraph4j.state.AgentState;
import org.bsc.langgraph4j.state.Channel;
import org.bsc.langgraph4j.state.Channels;
import java.io.Serializable;
import java.util.ArrayList;
import java.util.Map;
import java.util.Optional;
record RagDocument(String content, String source) implements Serializable {}
class CorrectiveRagState extends AgentState {
static final String QUESTION = "question";
static final String DOCUMENTS = "documents";
static final String WEB_SEARCH = "web_search";
static final String GENERATION = "generation";
static final Map<String, Channel<?>> SCHEMA = Map.of(
QUESTION, Channels.base(() -> ""),
DOCUMENTS, Channels.base(ArrayList::new),
WEB_SEARCH, Channels.base(() -> false),
GENERATION, Channels.base(() -> "")
);
CorrectiveRagState(Map<String, Object> initData) { super(initData); }
String question() { return this.<String>value(QUESTION).orElse(""); }
@SuppressWarnings("unchecked")
List<RagDocument> documents() {
return this.<List<RagDocument>>value(DOCUMENTS).orElse(List.of());
}
boolean webSearch() { return this.<Boolean>value(WEB_SEARCH).orElse(false); }
Optional<String> generation() { return value(GENERATION); }
}
Configure grading, rewriting, and generation¶
The relevance grader uses structured output. Any value other than yes is treated conservatively as irrelevant.
In [6]:
Copied!
import dev.langchain4j.model.openai.OpenAiChatModel;
import dev.langchain4j.model.output.structured.Description;
import dev.langchain4j.service.AiServices;
import dev.langchain4j.service.SystemMessage;
import dev.langchain4j.web.search.tavily.TavilyWebSearchEngine;
import java.time.Duration;
class GradeDocuments {
@Description("Relevance score: 'yes' if relevant, or 'no' if not relevant")
public String binaryScore;
}
interface DocumentGrader {
@SystemMessage("Grade whether the retrieved document is relevant to the question. "
+ "Return yes only when it contains information that can help answer the question; otherwise return no.")
GradeDocuments grade(@dev.langchain4j.service.UserMessage String prompt);
}
interface QuestionRewriter {
@SystemMessage("Rewrite the user question into a concise query optimized for web search. "
+ "Return only the rewritten query.")
String rewrite(@dev.langchain4j.service.UserMessage String question);
}
interface AnswerGenerator {
@SystemMessage("Answer the question using only the supplied context. "
+ "If the context is insufficient, say so. Keep the answer concise and cite source URLs when present.")
String answer(@dev.langchain4j.service.UserMessage String prompt);
}
var chatModelBuilder = OpenAiChatModel.builder()
.apiKey(openAiKey)
.modelName(chatModelName)
.temperature(0.0)
.timeout(Duration.ofMinutes(2))
.maxRetries(2);
if (openAiBaseUrl != null && !openAiBaseUrl.isBlank()) {
chatModelBuilder.baseUrl(openAiBaseUrl);
}
var chatModel = chatModelBuilder.build();
var documentGrader = AiServices.create(DocumentGrader.class, chatModel);
var questionRewriter = AiServices.create(QuestionRewriter.class, chatModel);
var answerGenerator = AiServices.create(AnswerGenerator.class, chatModel);
var webSearchEngine = TavilyWebSearchEngine.builder().apiKey(tavilyKey).build();
import dev.langchain4j.model.openai.OpenAiChatModel;
import dev.langchain4j.model.output.structured.Description;
import dev.langchain4j.service.AiServices;
import dev.langchain4j.service.SystemMessage;
import dev.langchain4j.web.search.tavily.TavilyWebSearchEngine;
import java.time.Duration;
class GradeDocuments {
@Description("Relevance score: 'yes' if relevant, or 'no' if not relevant")
public String binaryScore;
}
interface DocumentGrader {
@SystemMessage("Grade whether the retrieved document is relevant to the question. "
+ "Return yes only when it contains information that can help answer the question; otherwise return no.")
GradeDocuments grade(@dev.langchain4j.service.UserMessage String prompt);
}
interface QuestionRewriter {
@SystemMessage("Rewrite the user question into a concise query optimized for web search. "
+ "Return only the rewritten query.")
String rewrite(@dev.langchain4j.service.UserMessage String question);
}
interface AnswerGenerator {
@SystemMessage("Answer the question using only the supplied context. "
+ "If the context is insufficient, say so. Keep the answer concise and cite source URLs when present.")
String answer(@dev.langchain4j.service.UserMessage String prompt);
}
var chatModelBuilder = OpenAiChatModel.builder()
.apiKey(openAiKey)
.modelName(chatModelName)
.temperature(0.0)
.timeout(Duration.ofMinutes(2))
.maxRetries(2);
if (openAiBaseUrl != null && !openAiBaseUrl.isBlank()) {
chatModelBuilder.baseUrl(openAiBaseUrl);
}
var chatModel = chatModelBuilder.build();
var documentGrader = AiServices.create(DocumentGrader.class, chatModel);
var questionRewriter = AiServices.create(QuestionRewriter.class, chatModel);
var answerGenerator = AiServices.create(AnswerGenerator.class, chatModel);
var webSearchEngine = TavilyWebSearchEngine.builder().apiKey(tavilyKey).build();
Implement the nodes¶
Retrieve local context, grade every document independently, and preserve relevant documents for possible combination with web results.
In [7]:
Copied!
import dev.langchain4j.rag.query.Query;
import org.bsc.langgraph4j.action.EdgeAction;
import org.bsc.langgraph4j.action.NodeAction;
String localSource(TextSegment segment) {
var source = segment.metadata().getString(Document.URL);
return source == null || source.isBlank() ? "local-vector-store" : source;
}
NodeAction<CorrectiveRagState> retrieve = state -> {
var documents = retriever.retrieve(new Query(state.question())).stream()
.map(content -> new RagDocument(
content.textSegment().text(), localSource(content.textSegment())))
.toList();
return Map.of(
CorrectiveRagState.DOCUMENTS, new ArrayList<>(documents),
CorrectiveRagState.WEB_SEARCH, false
);
};
NodeAction<CorrectiveRagState> gradeDocuments = state -> {
var filtered = new ArrayList<RagDocument>();
// Empty retrieval is a failed retrieval and therefore needs correction.
var needsWebSearch = state.documents().isEmpty();
for (var document : state.documents()) {
var prompt = "Question: " + state.question()
+ "\n\nDocument:\n" + document.content();
var grade = documentGrader.grade(prompt);
if (grade != null && "yes".equalsIgnoreCase(grade.binaryScore)) {
filtered.add(document);
} else {
needsWebSearch = true;
}
}
return Map.of(
CorrectiveRagState.DOCUMENTS, filtered,
CorrectiveRagState.WEB_SEARCH, needsWebSearch
);
};
EdgeAction<CorrectiveRagState> decideToGenerate = state ->
state.webSearch() ? "transform_query" : "generate";
import dev.langchain4j.rag.query.Query;
import org.bsc.langgraph4j.action.EdgeAction;
import org.bsc.langgraph4j.action.NodeAction;
String localSource(TextSegment segment) {
var source = segment.metadata().getString(Document.URL);
return source == null || source.isBlank() ? "local-vector-store" : source;
}
NodeAction<CorrectiveRagState> retrieve = state -> {
var documents = retriever.retrieve(new Query(state.question())).stream()
.map(content -> new RagDocument(
content.textSegment().text(), localSource(content.textSegment())))
.toList();
return Map.of(
CorrectiveRagState.DOCUMENTS, new ArrayList<>(documents),
CorrectiveRagState.WEB_SEARCH, false
);
};
NodeAction<CorrectiveRagState> gradeDocuments = state -> {
var filtered = new ArrayList<RagDocument>();
// Empty retrieval is a failed retrieval and therefore needs correction.
var needsWebSearch = state.documents().isEmpty();
for (var document : state.documents()) {
var prompt = "Question: " + state.question()
+ "\n\nDocument:\n" + document.content();
var grade = documentGrader.grade(prompt);
if (grade != null && "yes".equalsIgnoreCase(grade.binaryScore)) {
filtered.add(document);
} else {
needsWebSearch = true;
}
}
return Map.of(
CorrectiveRagState.DOCUMENTS, filtered,
CorrectiveRagState.WEB_SEARCH, needsWebSearch
);
};
EdgeAction<CorrectiveRagState> decideToGenerate = state ->
state.webSearch() ? "transform_query" : "generate";
Rewrite only on the corrective path, then merge up to three Tavily results with the retained local context. A blank model rewrite falls back to the original query.
In [8]:
Copied!
import dev.langchain4j.web.search.WebSearchRequest;
NodeAction<CorrectiveRagState> transformQuery = state -> {
var rewritten = questionRewriter.rewrite(state.question());
// Keep a usable query if the model unexpectedly returns an empty rewrite.
var nextQuestion = rewritten == null || rewritten.isBlank()
? state.question() : rewritten;
return Map.of(CorrectiveRagState.QUESTION, nextQuestion);
};
String webResultText(String content, String snippet, String title) {
if (content != null && !content.isBlank()) return content;
if (snippet != null && !snippet.isBlank()) return snippet;
return title == null ? "" : title;
}
NodeAction<CorrectiveRagState> webSearch = state -> {
var merged = new ArrayList<>(state.documents());
webSearchEngine.search(WebSearchRequest.from(state.question(), 3))
.results().stream()
.map(result -> new RagDocument(
webResultText(result.content(), result.snippet(), result.title()),
result.url().toString()))
.forEach(merged::add);
return Map.of(CorrectiveRagState.DOCUMENTS, merged);
};
import dev.langchain4j.web.search.WebSearchRequest;
NodeAction<CorrectiveRagState> transformQuery = state -> {
var rewritten = questionRewriter.rewrite(state.question());
// Keep a usable query if the model unexpectedly returns an empty rewrite.
var nextQuestion = rewritten == null || rewritten.isBlank()
? state.question() : rewritten;
return Map.of(CorrectiveRagState.QUESTION, nextQuestion);
};
String webResultText(String content, String snippet, String title) {
if (content != null && !content.isBlank()) return content;
if (snippet != null && !snippet.isBlank()) return snippet;
return title == null ? "" : title;
}
NodeAction<CorrectiveRagState> webSearch = state -> {
var merged = new ArrayList<>(state.documents());
webSearchEngine.search(WebSearchRequest.from(state.question(), 3))
.results().stream()
.map(result -> new RagDocument(
webResultText(result.content(), result.snippet(), result.title()),
result.url().toString()))
.forEach(merged::add);
return Map.of(CorrectiveRagState.DOCUMENTS, merged);
};
In [9]:
Copied!
import java.util.stream.Collectors;
String formatContext(List<RagDocument> documents) {
return documents.stream()
.map(document -> "Source: " + document.source() + "\n" + document.content())
.collect(Collectors.joining("\n\n"));
}
NodeAction<CorrectiveRagState> generate = state -> {
var prompt = "Question: " + state.question()
+ "\n\nContext:\n" + formatContext(state.documents());
return Map.of(CorrectiveRagState.GENERATION, answerGenerator.answer(prompt));
};
import java.util.stream.Collectors;
String formatContext(List<RagDocument> documents) {
return documents.stream()
.map(document -> "Source: " + document.source() + "\n" + document.content())
.collect(Collectors.joining("\n\n"));
}
NodeAction<CorrectiveRagState> generate = state -> {
var prompt = "Question: " + state.question()
+ "\n\nContext:\n" + formatContext(state.documents());
return Map.of(CorrectiveRagState.GENERATION, answerGenerator.answer(prompt));
};
Assemble and run the graph¶
The graph has no correction loop: it performs at most one query rewrite and one web search before generation.
In [10]:
Copied!
import org.bsc.langgraph4j.StateGraph;
import static org.bsc.langgraph4j.StateGraph.END;
import static org.bsc.langgraph4j.StateGraph.START;
import static org.bsc.langgraph4j.action.AsyncEdgeAction.edge_async;
import static org.bsc.langgraph4j.action.AsyncNodeAction.node_async;
var graph = new StateGraph<>(CorrectiveRagState.SCHEMA, CorrectiveRagState::new)
.addNode("retrieve", node_async(retrieve))
.addNode("grade_documents", node_async(gradeDocuments))
.addNode("transform_query", node_async(transformQuery))
.addNode("web_search", node_async(webSearch))
.addNode("generate", node_async(generate))
.addEdge(START, "retrieve")
.addEdge("retrieve", "grade_documents")
.addConditionalEdges("grade_documents", edge_async(decideToGenerate), Map.of(
"transform_query", "transform_query",
"generate", "generate"
))
.addEdge("transform_query", "web_search")
.addEdge("web_search", "generate")
.addEdge("generate", END)
.compile();
import org.bsc.langgraph4j.StateGraph;
import static org.bsc.langgraph4j.StateGraph.END;
import static org.bsc.langgraph4j.StateGraph.START;
import static org.bsc.langgraph4j.action.AsyncEdgeAction.edge_async;
import static org.bsc.langgraph4j.action.AsyncNodeAction.node_async;
var graph = new StateGraph<>(CorrectiveRagState.SCHEMA, CorrectiveRagState::new)
.addNode("retrieve", node_async(retrieve))
.addNode("grade_documents", node_async(gradeDocuments))
.addNode("transform_query", node_async(transformQuery))
.addNode("web_search", node_async(webSearch))
.addNode("generate", node_async(generate))
.addEdge(START, "retrieve")
.addEdge("retrieve", "grade_documents")
.addConditionalEdges("grade_documents", edge_async(decideToGenerate), Map.of(
"transform_query", "transform_query",
"generate", "generate"
))
.addEdge("transform_query", "web_search")
.addEdge("web_search", "generate")
.addEdge("generate", END)
.compile();
In [11]:
Copied!
var result = graph.invoke(Map.of(
CorrectiveRagState.QUESTION,
"What are the main components of an LLM-powered autonomous agent system?"
)).orElseThrow();
System.out.println("Final query: " + result.question());
System.out.println("Web search used: " + result.webSearch());
System.out.println(result.generation().orElse("No answer generated"));
var result = graph.invoke(Map.of(
CorrectiveRagState.QUESTION,
"What are the main components of an LLM-powered autonomous agent system?"
)).orElseThrow();
System.out.println("Final query: " + result.question());
System.out.println("Web search used: " + result.webSearch());
System.out.println(result.generation().orElse("No answer generated"));
Final query: Main components of LLM-powered autonomous agent system Web search used: true The main components of an LLM-powered autonomous agent system include: 1. **Language Model Integration** - The core component that serves as the agent's brain. 2. **Decision-Making Modules** - For autonomous decision-making. 3. **Memory Module** - Allows the agent to store and recall past interactions and data. 4. **Execution Interfaces** - For carrying out tasks. 5. **Continuous Learning Capabilities** - To improve over time through learning from interactions. Additionally, planning and tool use are also highlighted as important components in some contexts. For more details, you can refer to the sources: - [Lilian Weng](https://lilianweng.github.io/posts/2023-06-23-agent/) - [EMA AI](https://www.ema.ai/additional-blogs/addition-blogs/introduction-to-autonomous-llm-powered-agents) - [K2View](https://www.k2view.com/blog/llm-powered-autonomous-agents) - [Truera](https://truera.com/ai-quality-education/generative-ai-agents/what-are-llm-powered-autonomous-agents)
Tests¶
CorrectiveRagStubTest covers the workflow without API keys or network access. CorrectiveRagITest exercises the same topology with the configured OpenAI-compatible service and Tavily and is excluded from default Surefire runs.