Comparing LLM4S with other JVM libraries
This page compares LLM4S with LangChain4j, Spring AI, Koog and Embabel, so that you can decide quickly whether it fits your project. It states facts and trade-offs, not rankings.
Last verified: 2026-10-11. The statements about LLM4S describe the main branch on that date; the latest release is
v0.4.1, and the module split described in the v1 scope ships with 0.5.0. Every statement about
another project comes from that project’s own documentation, fetched on that date and listed under
Sources. A cell that says not verified means it was not checked, not that the feature is absent. The
other projects change quickly: re-check anything that matters to you before relying on it, and expect this page to be
refreshed at each LLM4S release.
- At a glance, by audience
- Capability table
- What LLM4S has not got yet
- When not to choose LLM4S
- When LLM4S fits
- Sources
At a glance, by audience
- Scala developers. LLM4S is written for Scala 3 only (troubleshooting). It reports
errors as
Result[A]values, and its tools, agents, RAG and memory are Scala APIs. The other four projects in the table are Java or Kotlin libraries (see the first row); a Scala program can call them, but their documented APIs are written for those languages. - Java developers. LLM4S has a Java facade (
llm4s-java-api), but it is incomplete today.JLlmClientofferscompleteandcompletion, andJAgentruns, continues, streams the events of, resumes and recovers an agent run; defining tools, streaming tokens fromJLlmClientand structured output from Java are open issues (#1484, #1485, #1486). A Java team that wants a complete Java API today should read the table below carefully. - Spring shops. LLM4S has a Spring Boot starter (
llm4s-spring-boot-starter), but it wires three providers (OpenAI, Anthropic and Ollama; #1467 tracks the rest). Spring AI is the Spring project’s own library, and LangChain4j publishes Spring Boot starters too.
Capability table
| Capability | LLM4S (main) |
LangChain4j | Spring AI | Koog | Embabel |
|---|---|---|---|---|---|
| Language and API style | Scala 3 only; Result[A] errors; Java facade incomplete |
Java; AI Services are interfaces the library implements (annotated @AiService in Spring Boot) [L1][L8] |
Java; fluent ChatClient [S6] |
Kotlin DSL and fluent builder-style Java APIs [K1] | Written in Kotlin, usable from Java [E1] |
| Typed structured output | completeStructured[A] with a hand-written ObjectSchema[A] and a uPickle reader; the schema is sent as a JSON-Schema response format; Anthropic gets a best-effort prompt instruction that is not schema-enforced (its Scaladoc). No schema derivation yet (#1472) |
Return a POJO or record from an AI Service method; JSON Schema for the providers it lists, prompting otherwise [L1][L7] | BeanOutputConverter derives a JSON Schema from a Java class; entity() on ChatClient [S4][S6] |
not verified | not verified |
| Tools | ToolBuilder with a hand-written parameter schema; built-in calculator, date and time, UUID, JSON, file, HTTP, shell and web-search tools (guide); no derivation from types yet (#1473) |
AI Services can be configured with tools, methods annotated @Tool, that the model can call [L1] |
@Tool on a method, or MethodToolCallback and FunctionToolCallback; the input schema is generated from the method’s parameters [S7]; tool calls are observed [S5] |
not verified | not verified |
| Agents | Agent runtime on a typed graph runtime, handoffs (guide) | langchain4j-agentic, marked experimental: sequential, loop, parallel and conditional workflows and a supervisor agent [L10]; an example of an agent with memory, tools and RAG [L2] |
A guide to agentic patterns (chain, parallelization, routing, orchestrator-workers, evaluator-optimizer) built on ChatClient [S10]; the Advisors API intercepts and enhances model interactions [S6] |
Basic, graph-based, functional and planner (beta) agents [K1] | Goal Oriented Action Planning by default; Utility AI supported [E1] |
| Guardrails | Input and output guardrails, including LLM-as-judge ones, and composition (guide) | Input and output guardrails on AI Services, marked experimental [L11] | not verified | not verified | not verified |
| Memory | llm4s-memory: in-memory, SQLite and embedding-backed stores; a Postgres store in the separate llm4s-memory-postgres artifact (guide) |
Chat memory, shared or per user through ChatMemoryProvider [L1] |
ChatMemory, with MessageWindowChatMemory as its built-in implementation, backed by ChatMemoryRepository implementations, including in-memory, JDBC, Cassandra, Neo4j, MongoDB and Redis [S6] |
not verified | not verified |
| RAG and stores | Vector stores: SQLite, PostgreSQL (pgvector), Qdrant; keyword indexes: SQLite FTS5, PostgreSQL full-text; hybrid fusion and Cohere reranking (guide) | ContentRetriever and RetrievalAugmentor [L1]; a long table of embedding stores [L4] |
A list of vector store implementations [S2] | not verified | not verified |
| Model providers | 15 chat providers on main: OpenAI, Azure OpenAI, Anthropic, Gemini, Vertex AI, Ollama, Mistral, Cohere, DeepSeek, Z.ai, OpenRouter, Requesty, Amazon Bedrock, IBM watsonx.ai and a generic OpenAI-compatible one; 5 embedding providers (providers) |
A table of supported language models [L3] | A comparison table of supported chat models [S8] | not verified | not verified |
| MCP | Client (stdio, SSE and Streamable HTTP transports) and a server class in llm4s-mcp; no user guide yet (#1442) |
Client with Streamable HTTP, stdio, WebSocket and Docker stdio [L5]; a stdio server in the langchain4j-community-mcp-server module [L9] |
Client and server; stdio, Streamable HTTP, stateless Streamable HTTP and SSE; Spring Boot starters [S3] | MCP and A2A integrations, both marked beta [K1] | not verified |
| Observability | Langfuse, OpenTelemetry and Prometheus modules, a trace collector and cost tracking (guide) | Listeners and events; Micrometer listeners; OpenTelemetry semantic conventions; Arize integrations [L6] | Micrometer metrics and tracing for ChatClient, ChatModel, EmbeddingModel, ImageModel and VectorStore [S5] |
OpenTelemetry is listed among its integrations [K1] | not verified |
| Testing and evaluation | RAGAS-style metrics and a benchmark harness for RAG (guide); LLM-as-judge guardrails; no published test kit for application code (#1475); llm4s-provider-testkit is for provider authors |
not verified | Evaluator, RelevancyEvaluator and FactCheckingEvaluator [S1] |
not verified | not verified |
| Frameworks and runtimes | Spring Boot starter (three providers), cats-effect (guide), ZIO (guide), Pekko Streams (guide), a Kotlin module, a Java facade (incomplete) | Spring Boot 3 and 4 starters and @AiService; example projects for Spring Boot, Quarkus, Helidon, Payara Micro, WildFly and Jakarta EE/MicroProfile [L2][L8] |
Spring AI 2.0.x supports Spring Boot 4.0.x and 4.1.x [S9]; Spring Boot starters, for example for its MCP client and server [S3] | Spring Boot and Ktor integrations (both beta) [K1] | Built on Spring [E1] |
| Java baseline | JDK 21: CI runs only JDK 21, and the agent runtime depends on Ox, whose jars are JDK 21 bytecode (#1493) | Its Spring Boot integration page states a minimum Java version, quoted in the note below the table [L8] | not verified | not verified | not verified |
| Maturity | Pre-1.0: latest release v0.4.1; stability tiers (Frozen, Beta, Experimental) in the v1 scope | Latest release 1.22.0, 2026-10-08 [R1]; some modules, such as agents and guardrails, are marked experimental [L10][L11] | Latest GA release 2.0.1, 2026-08-21 [R2] | Latest release 1.3.0, 2026-09-24 [R3]; its Spring Boot, Ktor and planner agents are marked beta [K1] | Latest release 1.5.3, 2026-10-05 [R4] |
Note on the Java baseline row: LangChain4j’s Spring Boot integration page says “LangChain4j Spring Boot integration requires Java 17” [L8]. This is a statement about that project, quoted as fetched on 2026-10-11.
What LLM4S has not got yet
Plainly, as of the date above:
- Schema and tool derivation. Structured-output schemas and tool parameter schemas are written by hand, and a mismatch with the handler shows up at run time (#1472, #1473).
- A complete Java API. See the audience note above.
- A test kit for application code (#1475). Tests in this repository use scripted clients like the ones in the migration guide’s spec.
- A smaller integration catalogue. Two vector-store modules that other projects offer, Elasticsearch/OpenSearch and Redis, are open issues (#1469, #1470). The Spring Boot starter covers three providers (#1467). There is no Quarkus guide yet (#1468).
- No bill of materials to align artifact versions (#1462).
- No Scala 2.13 artifact. The artifacts are Scala 3 only.
- A stable 1.0 API. The split artifacts are unpublished until 0.5.0, and the compatibility promise starts at 1.0.
When not to choose LLM4S
- Your application is Java and needs a complete Java API now: defining tools, streaming tokens from a plain completion
and structured output from Java are not there on
mainyet. - You are a Spring team that wants the widest set of ready-made integrations, or you need a vector store or model provider that LLM4S does not have a module for. The two projects whose pages list their stores, LangChain4j and Spring AI, each list far more than the three vector stores LLM4S has.
- You need to stay on a JDK older than 21.
- You need a stable, frozen API today. LLM4S is pre-1.0.
When LLM4S fits
These are the things its code does today, not claims about quality:
- You write Scala 3 and want errors as
Result[A]values with typed error classes (error handling) instead of exceptions. - You want a provider layer that is a small published interface (
ProviderDescriptor), so that adding a provider is adding a dependency (writing a provider). - You want guardrails (input and output validators, including LLM-as-judge ones) and agent handoffs in the same library as the client, usable on an agent run (guardrails).
- You use cats-effect or ZIO and want
AgentIOorAgentZ(cats-effect, ZIO).
Sources
All fetched on 2026-10-11, against LangChain4j 1.22.0, Spring AI 2.0.1, Koog 1.3.0 and Embabel 1.5.3. The column shows the fact taken from each page; nothing else on these pages is relied on.
| Tag | Page | Fact used |
|---|---|---|
| L1 | LangChain4j: AI Services | Return-type structured output; tools; chat memory (shared or per user); ContentRetriever for RAG |
| L2 | langchain4j-examples | Example projects for Spring Boot, Quarkus, Helidon, Payara Micro, WildFly, Jakarta EE/MicroProfile; an agent example with memory, tools and RAG |
| L3 | LangChain4j: language models | A comparison table of supported language models |
| L4 | LangChain4j: embedding stores | A comparison table of supported embedding stores |
| L5 | LangChain4j: MCP | MCP client; Streamable HTTP, stdio, WebSocket (not standardized) and Docker stdio; stdio server in LangChain4j Community |
| L6 | LangChain4j: observability | Listeners and events; Micrometer listeners; OpenTelemetry Generative AI Semantic Conventions; Arize Phoenix and AX |
| L7 | LangChain4j: structured outputs | JSON Schema for Amazon Bedrock, Azure OpenAI, Google AI Gemini, Mistral, Ollama and OpenAI; prompting otherwise; POJOs and records |
| L8 | LangChain4j: Spring Boot integration | Starters for Spring Boot 3 (3.5+) and 4 (4.0+); @AiService; its stated Java minimum (see the note below the capability table) |
| L9 | LangChain4j: building a Java MCP server | A stdio MCP server in the community module langchain4j-community-mcp-server |
| L10 | LangChain4j: agents and agentic AI | langchain4j-agentic, experimental; sequential, loop, parallel, conditional workflows; supervisor agent |
| L11 | LangChain4j: guardrails | Input and output guardrails, only on AI Services; experimental |
| S1 | Spring AI: evaluation testing | Evaluator, RelevancyEvaluator, FactCheckingEvaluator |
| S2 | Spring AI: vector databases | A list of vector store implementations |
| S3 | Spring AI: MCP overview | Client and server; stdio, Streamable HTTP, stateless Streamable HTTP, SSE; Spring Boot starters |
| S4 | Spring AI: structured output converters | BeanOutputConverter derives a JSON Schema from a Java class |
| S5 | Spring AI: observability | Micrometer metrics and tracing for ChatClient, ChatModel, EmbeddingModel, ImageModel, VectorStore; spring.ai.tool observations |
| S6 | Spring AI: ChatClient | Fluent API; entity(); Advisors API; ChatMemory with MessageWindowChatMemory as the built-in implementation, and its ChatMemoryRepository implementations |
| S7 | Spring AI: tool calling | @Tool, MethodToolCallback, FunctionToolCallback; input schema generated from the method’s parameters |
| S8 | Spring AI: chat models comparison | A comparison table of supported chat models |
| S9 | Spring AI: getting started | Spring AI 2.0.x supports Spring Boot 4.0.x and 4.1.x |
| S10 | Spring AI: building effective agents | Chain, parallelization, routing, orchestrator-workers and evaluator-optimizer patterns |
| K1 | Koog documentation | JetBrains JVM agent framework; Kotlin DSL and fluent builder-style Java APIs; basic, graph, functional and planner (beta) agents; Spring Boot and Ktor (beta); OpenTelemetry; MCP and A2A (beta) |
| E1 | embabel-agent | Agent framework for the JVM; written in Kotlin with a natural Java usage model; GOAP by default and Utility AI; built on Spring |
| R1 | LangChain4j releases | Latest release 1.22.0, published 2026-10-08 |
| R2 | Spring AI releases | Latest GA release v2.0.1, published 2026-08-21 |
| R3 | Koog releases | Latest release 1.3.0, published 2026-09-24 |
| R4 | Embabel releases | Latest release v1.5.3, published 2026-10-05 |
If you spot a statement here that is out of date or unfair to another project, please open an issue or a pull request.