Docs & rules
langchain4j/langchain4jGitHub
This is a public, read-only report. Striff reads this repository's docs, turns each sentence that makes a claim about the code into a rule, and checks the rule against the code on the default branch. How this works
1+ stale names in these docs.
main7d5a0bcLast read Oct 3
Is this yours? Install to manage it
Once installed, Striff checks every pull request.
Documents
Rules
26 docs
langchain4j/langchain4jOpen repository
Stale names in these docs1
Each sentence below has a stale name: a type the default branch doesn't declare, or declares somewhere else. Edit the doc so it matches the code, or bring the type back.
- Gone
dev.langchain4j.web.search.searchapi.SearchApiEngineimport dev.langchain4j.web.search.searchapi.SearchApiEngine;
dev.langchain4j.web.search.searchapideclares nothing by this name. Closest names there:SearchApiWebSearchEngine,SearchApiWebSearchEngineBuilder,SearchApiClient.
langchain4j/langchain4j — documented rules
345 of 345 rules, printed October 14, 2026.
| The sentence in your docs | |||
|---|---|---|---|
| Read Oct 3 | AnthropicBatchChatModel implements the core BatchChatModel interface (submit, retrieve, cancel, list). | AnthropicBatchChatModel implements BatchChatModel | Holds |
BatchChatModel has submit | Holds | ||
BatchChatModel has retrieve | Holds | ||
BatchChatModel has cancel | Holds | ||
BatchChatModel has list | Holds | ||
| Read Oct 3 | The results will be available in AiMessage.attributes() under the key "server_tool_results": | AiMessage has an attributes method | Holds |
| Read Oct 3 | The same skills(...) parameter is available on AnthropicStreamingChatModel. | AnthropicStreamingChatModel has a skills method | Holds |
| Read Oct 3 | AnthropicChatModel and AnthropicStreamingChatModel return AnthropicTokenUsage in the response, which contains cacheCreationInputTokens and cacheReadInputTokens. | AnthropicTokenUsage has cacheCreationInputTokens | Holds |
AnthropicTokenUsage has cacheReadInputTokens | Holds | ||
| Read Oct 3 | AiMessage and ToolExecutionResultMessage carry an immutable attributes map, so set it via toBuilder(). | ToolExecutionResultMessage has attributes | Holds |
AiMessage has a toBuilder method | Holds | ||
ToolExecutionResultMessage has a toBuilder method | Holds | ||
| Read Oct 3 | The usual chat(...) methods are available: | GoogleAiGeminiChatModel has a chat method | Holds |
| Read Oct 3 | The latter additionally exposes Gemini-only options such as aspectRatio and imageSize: | GoogleAiGeminiChatRequestParameters has aspectRatio | Holds |
GoogleAiGeminiChatRequestParameters has imageSize | Holds | ||
| Read Oct 3 | To read it, cast ChatResponse.metadata() to GoogleAiGeminiChatResponseMetadata: | ChatResponse has a metadata method | Holds |
| Read Oct 3 | Method safetyRatings(), Meaning How the generated content was rated, one entry per harm category Empty if none | GoogleAiGeminiChatResponseMetadata has a safetyRatings method | Holds |
| Read Oct 3 | Method promptSafetyRatings(), Meaning How your prompt was rated Empty if none | GoogleAiGeminiChatResponseMetadata has a promptSafetyRatings method | Holds |
| Read Oct 3 | Method blockReason(), Meaning Why Gemini refused the prompt outright, or null if it did not | GoogleAiGeminiChatResponseMetadata has a blockReason method | Holds |
| Read Oct 3 | GeminiSafetyRating.category() and probability() are plain Strings holding the raw API values, so harm categories that Google introduces later are passed through as-is instead of breaking your application. | GeminiSafetyRating has a category method | Holds |
GeminiSafetyRating has a probability method | Holds | ||
| Read Oct 3 | [1] - The JsonSchema can be generated automatically from your class using JsonSchemas.jsonSchemaFrom() helper method. | JsonSchemas has a jsonSchemaFrom method | Holds |
| Read Oct 3 | 2 rules from this sentence | GeminiThinkingConfig has includeThoughts | Holds |
GeminiThinkingConfig has thinkingBudget | Holds | ||
| Read Oct 3 | controls whether to return thinking (if available) inside AiMessage.thinking() and whether to invoke StreamingChatResponseHandler.onPartialThinking() and TokenStream.onPartialThinking() callbacks when using GoogleAiGeminiStreamingChatModel. | AiMessage has a thinking method | Holds |
StreamingChatResponseHandler has an onPartialThinking method | Holds | ||
TokenStream has an onPartialThinking method | Holds | ||
| Read Oct 3 | If enabled, tinking signatures will also be stored and returned inside the AiMessage.attributes(). | AiMessage has an attributes method | Holds |
| Read Oct 3 | :::note Please note that when returnThinking is not set (is null) and thinkingConfig is set, thinking text will be prepended to the actual response inside the AiMessage.text() field and StreamingChatResponseHandler.onPartialResponse() will be invoked instead of StreamingChatResponseHandler.onPartialThinking(). | AiMessage has a text method | Holds |
StreamingChatResponseHandler has an onPartialResponse method | Holds | ||
| Read Oct 3 | createCache has three forms for expiration: with no expiration argument it uses the API default (currently 1 hour); | GeminiCaches has createCache | Holds |
| Read Oct 3 | The BatchState enum represents the possible states of a batch job: | BatchState is an enumeration | Holds |
| Read Oct 3 | Use BatchState.isTerminal() on BatchResponse.state() to detect when polling can stop. | BatchState has an isTerminal method | Holds |
| Read Oct 3 | As well as the ChatResponse chat(ChatRequest req) method: | GoogleGenAiChatModel has a chat method | Holds |
| Read Oct 3 | The response must be handled by a StreamingChatResponseHandler. | GoogleGenAiStreamingChatModel depends on StreamingChatResponseHandler | Holds |
| Read Oct 3 | Instead of creating the cache out-of-band, you can create and manage it directly from LangChain4j with GoogleGenAiCaches, which wraps the SDK's cache lifecycle (create / get / list / update TTL / delete). The messages are cached using the same GoogleGenAiContentMapper the chat models use, so you stay in the LangChain4j ChatMessage domain. | GoogleGenAiCaches depends on GoogleGenAiContentMapper | Holds |
| Read Oct 3 | They arrive as inlineData parts, and GoogleGenAiChatModel maps them into the AiMessage attributes under GENERATED_IMAGES_KEY, which is what AiMessage.images() reads: | AiMessage has an images method | Holds |
| Read Oct 3 | The Mistral-specific options ( safePrompt, randomSeed, sendThinking and returnThinking) can also be set per request via MistralAiChatRequestParameters, overriding the values configured on the model builder. | MistralAiChatRequestParameters has safePrompt | Holds |
MistralAiChatRequestParameters has randomSeed | Holds | ||
MistralAiChatRequestParameters has sendThinking | Holds | ||
MistralAiChatRequestParameters has returnThinking | Holds | ||
| Read Oct 3 | when enabled, reasoning text produced by the model will be parsed from the API response and stored in AiMessage.thinking(). | AiMessage has a thinking method | Holds |
| Read Oct 3 | For streaming, StreamingChatResponseHandler.onPartialThinking() and TokenStream.onPartialThinking() callbacks will also be invoked. | StreamingChatResponseHandler has an onPartialThinking method | Holds |
TokenStream has an onPartialThinking method | Holds | ||
| Read Oct 3 | MistralAiBatchChatModel implements the core BatchChatModel interface to process many chat requests asynchronously via the Mistral Batch API, at 50% of the standard per-token price. | MistralAiBatchChatModel implements BatchChatModel | Holds |
| Read Oct 3 | In the previous section, an OpenAiOfficialChatModel object was created, which implements the ChatModel interface. | OpenAiOfficialChatModel implements ChatModel | Holds |
| Read Oct 3 | OpenAiOfficialPromptCacheBreakpoint.mark() returns a marked copy of the message, leaving the original untouched: | OpenAiOfficialPromptCacheBreakpoint has a mark method | Holds |
| Read Oct 3 | OpenAiOfficialResponsesChatRequestParameters extends DefaultChatRequestParameters with Responses API-specific fields: previousResponseId, maxToolCalls, parallelToolCalls, topLogprobs, truncation, include, serviceTier, safetyIdentifier, promptCacheKey, promptCacheRetention, promptCacheOptions, reasoningEffort, reasoningSummary, textVerbosity, streamIncludeObfuscation, store, strictTools, strictJsonSchema. | OpenAiOfficialResponsesChatRequestParameters extends DefaultChatRequestParameters | Holds |
| Read Oct 3 | previousResponseId, maxToolCalls, parallelToolCalls, topLogprobs, truncation, include, serviceTier, safetyIdentifier, promptCacheKey, promptCacheRetention, promptCacheOptions, reasoningEffort, reasoningSummary, textVerbosity, streamIncludeObfuscation, store, strictTools, strictJsonSchema. | OpenAiOfficialResponsesChatRequestParameters has previousResponseId | Holds |
OpenAiOfficialResponsesChatRequestParameters has maxToolCalls | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has parallelToolCalls | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has topLogprobs | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has truncation | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has include | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has serviceTier | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has safetyIdentifier | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has promptCacheKey | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has promptCacheRetention | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has promptCacheOptions | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has reasoningEffort | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has reasoningSummary | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has textVerbosity | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has streamIncludeObfuscation | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has store | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has strictTools | Holds | ||
OpenAiOfficialResponsesChatRequestParameters has strictJsonSchema | Holds | ||
| Read Oct 3 | When reasoningSummary is set for OpenAiOfficialResponsesStreamingChatModel, the StreamingChatResponseHandler.onPartialThinking() callback will be invoked as reasoning summary tokens are streamed: | StreamingChatResponseHandler has an onPartialThinking method | Holds |
| Read Oct 3 | The reasoning summary in AiMessage.thinking() is informational and does not need to be sent back in follow-up requests — OpenAI discards it between turns. | AiMessage has a thinking method | Holds |
| Read Oct 3 | This is automatically stored in AiMessage.attributes() under the key "encrypted_reasoning". | AiMessage has an attributes method | Holds |
| Read Oct 3 | OpenAiOfficialBatchChatModel implements the core BatchChatModel interface to process many chat requests asynchronously via the OpenAI Batch API, at 50% of the standard per-token price. | OpenAiOfficialBatchChatModel implements BatchChatModel | Holds |
| Read Oct 3 | It returns the generated audio as raw bytes wrapped in an Audio object. | OpenAiTextToSpeechModel depends on Audio | Holds |
| Read Oct 3 | OpenAiPromptCacheOptions.implicit() and OpenAiPromptCacheOptions.explicit() are shorthands for options that only set the mode. | OpenAiPromptCacheOptions has an implicit method | Holds |
OpenAiPromptCacheOptions has an explicit method | Holds | ||
| Read Oct 3 | OpenAiPromptCacheBreakpoint.mark() returns a marked copy of the message, leaving the original untouched: | OpenAiPromptCacheBreakpoint has a mark method | Holds |
| Read Oct 3 | OpenAiResponsesChatRequestParameters extends DefaultChatRequestParameters with Responses API-specific fields: previousResponseId, maxToolCalls, parallelToolCalls, topLogprobs, truncation, include, serviceTier, safetyIdentifier, promptCacheKey, promptCacheRetention, promptCacheOptions, reasoningEffort, reasoningSummary, textVerbosity, streamIncludeObfuscation, store, strictTools, strictJsonSchema. | OpenAiResponsesChatRequestParameters extends DefaultChatRequestParameters | Holds |
| Read Oct 3 | previousResponseId, maxToolCalls, parallelToolCalls, topLogprobs, truncation, include, serviceTier, safetyIdentifier, promptCacheKey, promptCacheRetention, promptCacheOptions, reasoningEffort, reasoningSummary, textVerbosity, streamIncludeObfuscation, store, strictTools, strictJsonSchema. | OpenAiResponsesChatRequestParameters has previousResponseId | Holds |
OpenAiResponsesChatRequestParameters has maxToolCalls | Holds | ||
OpenAiResponsesChatRequestParameters has parallelToolCalls | Holds | ||
OpenAiResponsesChatRequestParameters has topLogprobs | Holds | ||
OpenAiResponsesChatRequestParameters has truncation | Holds | ||
OpenAiResponsesChatRequestParameters has include | Holds | ||
OpenAiResponsesChatRequestParameters has serviceTier | Holds | ||
OpenAiResponsesChatRequestParameters has safetyIdentifier | Holds | ||
OpenAiResponsesChatRequestParameters has promptCacheKey | Holds | ||
OpenAiResponsesChatRequestParameters has promptCacheRetention | Holds | ||
OpenAiResponsesChatRequestParameters has promptCacheOptions | Holds | ||
OpenAiResponsesChatRequestParameters has reasoningEffort | Holds | ||
OpenAiResponsesChatRequestParameters has reasoningSummary | Holds | ||
OpenAiResponsesChatRequestParameters has textVerbosity | Holds | ||
OpenAiResponsesChatRequestParameters has streamIncludeObfuscation | Holds | ||
OpenAiResponsesChatRequestParameters has store | Holds | ||
OpenAiResponsesChatRequestParameters has strictTools | Holds | ||
OpenAiResponsesChatRequestParameters has strictJsonSchema | Holds | ||
| Read Oct 3 | When reasoningSummary is set for OpenAiResponsesStreamingChatModel, the StreamingChatResponseHandler.onPartialThinking() callback will be invoked as reasoning summary tokens are streamed: | StreamingChatResponseHandler has an onPartialThinking method | Holds |
| Read Oct 3 | The reasoning summary in AiMessage.thinking() is informational and does not need to be sent back in follow-up requests — OpenAI discards it between turns. | AiMessage has a thinking method | Holds |
| Read Oct 3 | This is automatically stored in AiMessage.attributes() under the key "encrypted_reasoning". | AiMessage has an attributes method | Holds |
| Read Oct 3 | The WatsonxChatModel class allows you to create an instance of the ChatModel interface fully encapsulated within LangChain4j. | WatsonxChatModel implements ChatModel | Holds |
| Read Oct 3 | Each watsonx.ai chat service has its own ChatRequestParameters implementation, exposing the parameters that service accepts and nothing else: Class WatsonxChatRequestParameters, Used by WatsonxChatModel , WatsonxStreamingChatModel , WatsonxDeploymentChatModel , WatsonxDeploymentStreamingChatModel | WatsonxChatRequestParameters implements ChatRequestParameters | Holds |
| Read Oct 3 | Parameter serviceTier(...), Description Service tier, one of AUTO , DEFAULT , FLEX , or PRIORITY | WatsonxGatewayChatRequestParameters has a serviceTier method | Holds |
| Read Oct 3 | Parameter reasoningEffort(...), Description Reasoning effort for reasoning models, one of LOW , MEDIUM , or HIGH | WatsonxGatewayChatRequestParameters has a reasoningEffort method | Holds |
| Read Oct 3 | Parameter router(...) / cache(...), Description Router configuration, including a prompt Cache | WatsonxGatewayChatRequestParameters has a router method | Holds |
| Read Oct 3 | Parameter modalities(...), Description Output modalities (e.g ["text"] ) | WatsonxGatewayChatRequestParameters has a modalities method | Holds |
| Read Oct 3 | Parameter store(...), Description Whether the provider should persist the request/response | WatsonxGatewayChatRequestParameters has a store method | Holds |
| Read Oct 3 | Parameter parallelToolCalls(...), Description Enable/disable parallel tool calls | WatsonxGatewayChatRequestParameters has a parallelToolCalls method | Holds |
| Read Oct 3 | Parameter user(...), Description End-user identifier forwarded to the provider | WatsonxGatewayChatRequestParameters has an user method | Holds |
| Read Oct 3 | Parameter metadata(...), Description Free-form metadata map forwarded to the provider | WatsonxGatewayChatRequestParameters has a metadata method | Holds |
| Read Oct 3 | Parameter logitBias(...) , logprobs(...) , topLogprobs(...) , seed(...), Description OpenAI-compatible sampling controls | WatsonxGatewayChatRequestParameters has a logitBias method | Holds |
WatsonxGatewayChatRequestParameters has a logprobs method | Holds | ||
WatsonxGatewayChatRequestParameters has a topLogprobs method | Holds | ||
WatsonxGatewayChatRequestParameters has a seed method | Holds | ||
| Read Oct 3 | It implements the LangChain4j ModelCatalog interface, allowing you to retrieve detailed information about each model. | WatsonxModelCatalog implements ModelCatalog | Holds |
| Read Oct 3 | It also implements the LangChain4j ModelCatalog interface. | WatsonxGatewayModelCatalog implements ModelCatalog | Holds |
| Read Oct 3 | The name() of every returned ModelDescription is the identifier to pass to WatsonxGatewayChatModel.modelName(...) and WatsonxGatewayStreamingChatModel.modelName(...). | ModelDescription has a name method | Holds |
| Read Oct 3 | ModelDescription displayName(), Gateway field same as name(), Notes the gateway has no separate label | ModelDescription has a displayName method | Holds |
| Read Oct 3 | ModelDescription description(), Gateway field description, Notes user-defined, null unless the administrator set it | ModelDescription has a description method | Holds |
| Read Oct 3 | ModelDescription owner(), Gateway field owned_by, Notes provider, e.g openai | ModelDescription has an owner method | Holds |
| Read Oct 3 | ModelDescription createdAt(), Gateway field created, Notes Unix timestamp of the gateway configuration , not of the model release | ModelDescription has a createdAt method | Holds |
| Read Oct 3 | ModelDescription type(), Gateway field -, Notes always ModelType.CHAT because the gateway does not expose model capabilities | ModelDescription has a type method | Holds |
| Read Oct 3 | ModelDescription maxInputTokens(), Gateway field metadata.context_window, Notes null when the administrator did not configure the metadata | ModelDescription has a maxInputTokens method | Holds |
| Read Oct 3 | ModelDescription maxOutputTokens(), Gateway field -, Notes always null , the gateway does not return it | ModelDescription has a maxOutputTokens method | Holds |
| Read Oct 3 | The WatsonxTokenCountEstimator implements the LangChain4j TokenCountEstimator interface by calling the watsonx.ai tokenization endpoint, so the count comes from the tokenizer of the model itself rather than from a local approximation. | WatsonxTokenCountEstimator implements TokenCountEstimator | Holds |
| Read Oct 3 | estimateTokenCountInMessage(...) also counts the thinking text and the > tool execution requests of an AiMessage. | WatsonxTokenCountEstimator has an estimateTokenCountInMessage method | Holds |
| Read Oct 3 | The WatsonxModerationModel provides a LangChain4j implementation of the ModerationModel interface using IBM watsonx.ai. | WatsonxModerationModel implements ModerationModel | Holds |
| Read Oct 3 | As you can see, there is a simple chat method that takes a String as input and returns a String as output, similar to LanguageModel. | ChatModel has a chat method | Holds |
| Read Oct 3 | ChatMessage is a base interface that represents a chat message. | ChatMessage is a contract | Holds |
| Read Oct 3 | One can also specify DetailLevel enum (with LOW/HIGH/AUTO options) to control how the model processes the image. | DetailLevel is an enumeration | Holds |
| Read Oct 3 | :::warning Breaking change for Kotlin users. ChatModel now declares a Java member chatAsync(ChatRequest): CompletableFuture<ChatResponse>. | ChatModel has a chatAsync method | Holds |
| Read Oct 3 | The DecisionModel API itself is part of langchain4j-core, which comes with every implementation: requests and questions are in dev.langchain4j.model.decision.request, answers in dev.langchain4j.model.decision.response and listeners in dev.langchain4j.model.decision.listener. | DecisionModel is in langchain4j-core | Holds |
| Read Oct 3 | A DecisionRequest contains the input (state) and the questions, each registered under a name of your choice: | DecisionRequest has input | Holds |
DecisionRequest has questions | Holds | ||
| Read Oct 3 | YesNoQuestion.of(text), ChoiceQuestion.of(text, options) and ScaleQuestion.of(text, levels). | YesNoQuestion has an of method | Holds |
ChoiceQuestion has an of method | Holds | ||
ScaleQuestion has an of method | Holds | ||
| Read Oct 3 | The response also carries the name of the model that produced the answers and the token usage: response.modelName() and response.tokenUsage(). | DecisionResponse has a modelName method | Holds |
DecisionResponse has a tokenUsage method | Holds | ||
| Read Oct 3 | A YesNoQuestion asks a yes/no question and is answered with a YesNoAnswer, whose probability() is the probability that the answer is "yes", from 0 to 1. | YesNoAnswer has a probability method | Holds |
| Read Oct 3 | isYes(threshold) turns it into a decision: | YesNoAnswer has an isYes method | Holds |
| Read Oct 3 | probabilityOf(option) and margin(): the probability of one option, and the difference between the two most likely options | ChoiceAnswer has a probabilityOf method | Holds |
ChoiceAnswer has a margin method | Holds | ||
| Read Oct 3 | Choice and scale answers carry the probability of each option or level, if the model reports them (otherwise probabilities() is empty). | ChoiceAnswer has a probabilities method | Holds |
ScaleAnswer has a probabilities method | Holds | ||
| Read Oct 3 | Choice and scale answers can also carry a confidence() value from 0 to 1. | ChoiceAnswer has a confidence method | Holds |
ScaleAnswer has a confidence method | Holds | ||
| Read Oct 3 | decideAsync(request) returns a CompletableFuture<DecisionResponse>. | DecisionModel has a decideAsync method | Holds |
| Read Oct 3 | Question and DecisionAnswer are interfaces, so an implementation can support additional question types with their own answer types. | Question is a contract | Holds |
DecisionAnswer is a contract | Holds | ||
| Read Oct 3 | Answers of such types can be read with response.answer(name, type), where type is the answer class defined by the implementation. | DecisionResponse has an answer method | Holds |
| Read Oct 3 | Component DecisionModelInputGuardrail , DecisionModelOutputGuardrail, Module langchain4j-guardrails | DecisionModelInputGuardrail is in langchain4j-guardrails | Holds |
DecisionModelOutputGuardrail is in langchain4j-guardrails | Holds | ||
| Read Oct 3 | Component DecisionScoringModel , DecisionModelQueryRouter , RoutingChatModel + DecisionModelChatModelRouter, Module langchain4j-core | DecisionScoringModel is in langchain4j-core | Holds |
DecisionModelQueryRouter is in langchain4j-core | Holds | ||
RoutingChatModel is in langchain4j-core | Holds | ||
DecisionModelChatModelRouter is in langchain4j-core | Holds | ||
| Read Oct 3 | DecisionScoringModel is a ScoringModel: the score of a segment is the probability that the answer to "Does the document help answer the query?" is "yes". | DecisionScoringModel implements ScoringModel | Holds |
| Read Oct 3 | Each segment is asked as a separate yes/no question, and the segments are scored in requests of up to 20 segments ( maxSegmentsPerRequest(...)), which are sent in parallel by scoreAsync(...). | DecisionScoringModel has a maxSegmentsPerRequest method | Holds |
DecisionScoringModel has a scoreAsync method | Holds | ||
| Read Oct 3 | The decision model receives the query and, when it comes from a conversation, the 2 previous messages ( maxMessages(3) by default), so that follow-up questions such as "and for contractors?" are understood. | DecisionModelQueryRouter has a maxMessages method | Holds |
| Read Oct 3 | fallbackStrategy(ROUTE_TO_ALL) retrieves from all sources instead, which favors answer quality. | DecisionModelQueryRouter has a fallbackStrategy method | Holds |
| Read Oct 3 | DecisionModelToolSearchStrategy is a tool search strategy: the LLM searches for tools when it needs them, and the decision model decides which tools match the search. | DecisionModelToolSearchStrategy implements ToolSearchStrategy | Holds |
| Read Oct 3 | DecisionModelFilteringToolProvider wraps a ToolProvider and passes on only the tools that are relevant to the conversation, before the first LLM call, so no tool search round trip is needed. | DecisionModelFilteringToolProvider depends on ToolProvider | Holds |
| Read Oct 3 | If the wrapped tool provider is dynamic ( isDynamic() returns true), the AI Service asks it for tools before each LLM call of the tool-calling loop, so the decision model is called each time as well, which adds its latency to each round. | ToolProvider has an isDynamic method | Holds |
| Read Oct 3 | Component DecisionModelChatModelRouter, Default behavior when the decision model fails the default route is used, a warning is logged, Configurable fallbackStrategy | DecisionModelChatModelRouter has fallbackStrategy | Holds |
| Read Oct 3 | Component DecisionModelFilteringToolProvider, Default behavior when the decision model fails all tools are passed on, a warning is logged, Configurable fallbackStrategy | DecisionModelFilteringToolProvider has fallbackStrategy | Holds |
| Read Oct 3 | All of these exceptions extend LangChain4jException. | InvalidDecisionResponseException extends LangChain4jException | Holds |
| Read Oct 3 | mean() is the probability-weighted average of the level indexes, so it can fall between two levels. | Scale has a mean method | Holds |
| Read Oct 3 | probabilityAtLeast(level) is useful to act on a level or anything above it, for example to page the on-call engineer when an incident is likely to be at least of high severity: | Scale has a probabilityAtLeast method | Holds |
| Read Oct 3 | serviceInterface() and method(): the service and the method that is invoked; | ThresholdContext has a serviceInterface method | Holds |
ThresholdContext has a method method | Holds | ||
| Read Oct 3 | Input guardrails are implemented by implementing the InputGuardrail interface. | InputGuardrail is a contract | Holds |
| Read Oct 3 | The InputGuardrail interface has two variants of the validate method, at least one of which needs to be implemented: | InputGuardrail has a validate method | Holds |
| Read Oct 3 | Outcome success with alternate result, Helper method on InputGuardrail successWith(String), Description - Similar to success except the user message is altered before proceeding to the next step (next guardrail in the chain or calling the LLM). - A rewrite made by a guardrail overrides any rewrite made by an earlier guardrail in the chain | InputGuardrail has a successWith method | Holds |
| Read Oct 3 | Outcome fatal, Helper method on InputGuardrail fatal(String) or fatal(String, Throwable), Description - The input is invalid and execution is halted with an InputGuardrailException . - The LLM is not called. - If a Throwable is passed, consumers can catch InputGuardrailException and check the cause It will be the Throwable passed here | InputGuardrail has a fatal method | Holds |
| Read Oct 3 | Similar to input guardrails, output guardrails are implemented by implementing the OutputGuardrail interface. | OutputGuardrail is a contract | Holds |
| Read Oct 3 | The OutputGuardrail interface has two variants of the validate method, at least one of which needs to be implemented: | OutputGuardrail has a validate method | Holds |
| Read Oct 3 | Outcome failure, Helper method on OutputGuardrail failure(String) or failure(String, Throwable), Description - The output is invalid but the next guardrails in the chain continue to be executed in order to accumulate all possible validation problems. - The validation failure is returned to the user as an OutputGuardrailException | OutputGuardrail has a failure method | Holds |
| Read Oct 3 | Outcome fatal, Helper method on OutputGuardrail fatal(String) or fatal(String, Throwable), Description The output is invalid and execution is halted with an OutputGuardrailException thrown to the caller | OutputGuardrail has a fatal method | Holds |
| Read Oct 3 | LangChain4j exposes all of this metadata through the map stored inside ToolSpecification.metadata() method. | ToolSpecification has a metadata method | Holds |
| Read Oct 3 | The title field that exists directly in the MCP tool definition is exposed under McpToolMetadataKeys.TITLE key in the metadata map to distinguish it from the title that is retrieved from annotations - that one is exposed under the McpToolMetadataKeys.ANNOTATION_TITLE key. | McpToolMetadataKeys has a TITLE field | Holds |
| Read Oct 3 | If the tool has icons, they are exposed under the McpToolMetadataKeys.ICONS key in the metadata map. | McpToolMetadataKeys has ICONS | Holds |
| Read Oct 3 | The entries of the _meta object are exposed through ToolExecutionResult.attributes(), using their original keys, with the JSON values converted into nested maps. | ToolExecutionResult has an attributes method | Holds |
| Read Oct 3 | To keep them, enable returnToolResultAttributes on the tool provider: | McpToolProvider has returnToolResultAttributes | Holds |
| Read Oct 3 | It is still delivered to McpClientListener.afterExecuteTool(), which receives the complete raw response. | McpClientListener has an afterExecuteTool method | Holds |
| Read Oct 3 | McpApplicationErrorException therefore implements ToolErrorVisibleToLlm: with a tool execution error handler that honors it, such as ToolExecutionErrorHandler.failInvocationUnlessVisibleToLlm(), the server's text is sent to the LLM and it can react to it. | McpApplicationErrorException implements ToolErrorVisibleToLlm | Holds |
| Read Oct 3 | with a tool execution error handler that honors it, such as ToolExecutionErrorHandler.failInvocationUnlessVisibleToLlm(), the server's text is sent to the LLM and it can react to it. | ToolExecutionErrorHandler has a failInvocationUnlessVisibleToLlm method | Holds |
| Read Oct 3 | The supplier receives an McpCallContext (nullable) that contains the message being sent and, when applicable, the InvocationContext of the AI service call that triggered it. | McpCallContext depends on InvocationContext | Holds |
| Read Oct 3 | If you want to change this behavior, there is an interface named dev.langchain4j.mcp.client.logging.McpLogMessageHandler that serves as a callback for received log messages. | McpLogMessageHandler is a contract | Holds |
| Read Oct 3 | The interface dev.langchain4j.mcp.client.McpClientListener serves as the base for listener implementations. | McpClientListener is a contract | Holds |
McpClientListener is in dev.langchain4j.mcp.client | Holds | ||
| Read Oct 3 | Each McpResourceContents object represents either a binary blob (McpBlobResourceContents) or text (McpTextResourceContents). | McpBlobResourceContents depends on McpResourceContents | Holds |
McpTextResourceContents depends on McpResourceContents | Holds | ||
| Read Oct 3 | If you set a McpResourcesAsToolsPresenter instance using the builder when building an McpToolProvider, the MCP tool provider will automatically add two synthetic tools to the result of its provideTools method, along the 'regular' tools that are supported by the backing MCP servers. One tool should serve for obtaining a list of resources and the other for obtaining a particular resource. LangChain4j provides a default implementation named DefaultMcpResourcesAsToolsPresenter that adds these two tools: | DefaultMcpResourcesAsToolsPresenter implements McpResourcesAsToolsPresenter | Holds |
| Read Oct 3 | You can use McpPromptMessage.toChatMessage() to convert it into a generic dev.langchain4j.data.message.ChatMessage from the LangChain4j core API. | McpPromptMessage has a toChatMessage method | Holds |
| Read Oct 3 | You can manually clear this cache by calling DefaultMcpClient.evictToolListCache(). | DefaultMcpClient has an evictToolListCache method | Holds |
| Read Oct 3 | Then, to search for MCP servers, use the registry.listServers(McpServerListRequest) method. | DefaultMcpRegistryClient has a listServers method | Holds |
| Read Oct 3 | The seam is there — the CompletableFutureAdapter and PublisherAdapter SPIs, discovered via ServiceLoader — but no adapter ships for them, so declaring such a return type currently fails. | CompletableFutureAdapter is a contract | Holds |
PublisherAdapter is a contract | Holds | ||
| Read Oct 3 | Layer Chat model, Blocking chat(ChatRequest), Non-blocking counterpart chatAsync(ChatRequest) , and chat(ChatRequest) returning a Publisher on StreamingChatModel | ChatModel has a chatAsync method | Holds |
StreamingChatModel has a chat method | Holds | ||
ChatModel has a chat method | Holds | ||
| Read Oct 3 | Layer Embedding model, Blocking embed(...), Non-blocking counterpart embedAsync(...) | EmbeddingModel has an embedAsync method | Holds |
EmbeddingModel has an embed method | Holds | ||
| Read Oct 3 | Layer Scoring model, Blocking scoreAll(...), Non-blocking counterpart scoreAsync(...) | ScoringModel has a scoreAsync method | Holds |
ScoringModel has a scoreAll method | Holds | ||
| Read Oct 3 | Layer Chat memory, Blocking add , messages , set, Non-blocking counterpart addAsync , messagesAsync , setAsync | ChatMemory has addAsync | Holds |
ChatMemory has messagesAsync | Holds | ||
ChatMemory has setAsync | Holds | ||
ChatMemory has add | Holds | ||
ChatMemory has messages | Holds | ||
ChatMemory has set | Holds | ||
| Read Oct 3 | Layer Chat memory store, Blocking getMessages , updateMessages , deleteMessages, Non-blocking counterpart getMessagesAsync , updateMessagesAsync , deleteMessagesAsync | ChatMemoryStore has getMessagesAsync | Holds |
ChatMemoryStore has updateMessagesAsync | Holds | ||
ChatMemoryStore has deleteMessagesAsync | Holds | ||
ChatMemoryStore has getMessages | Holds | ||
ChatMemoryStore has updateMessages | Holds | ||
ChatMemoryStore has deleteMessages | Holds | ||
| Read Oct 3 | Layer Guardrails, Blocking validate(...), Non-blocking counterpart validateAsync(...) | Guardrail has a validateAsync method | Holds |
Guardrail has a validate method | Holds | ||
| Read Oct 3 | Layer Tools, Blocking ToolExecutor.execute(...), Non-blocking counterpart ToolExecutor.executeAsync(...) | ToolExecutor has an executeAsync method | Holds |
ToolExecutor has an execute method | Holds | ||
| Read Oct 3 | Layer RAG, Blocking RetrievalAugmentor.augment(...) , retrievers, routers, aggregators, query transformers, Non-blocking counterpart augmentAsync(...) , retrieveAsync(...) , routeAsync(...) , aggregateAsync(...) , transformAsync(...) | RetrievalAugmentor has an augmentAsync method | Holds |
ContentRetriever has a retrieveAsync method | Holds | ||
QueryRouter has a routeAsync method | Holds | ||
ContentAggregator has an aggregateAsync method | Holds | ||
QueryTransformer has a transformAsync method | Holds | ||
RetrievalAugmentor has an augment method | Holds | ||
| Read Oct 3 | Layer Embedding store, Blocking search(...), Non-blocking counterpart searchAsync(...) | EmbeddingStore has a searchAsync method | Holds |
EmbeddingStore has a search method | Holds | ||
| Read Oct 3 | Layer Web search, Blocking search(...), Non-blocking counterpart searchAsync(...) | WebSearchEngine has a search method | Holds |
WebSearchEngine has a searchAsync method | Holds | ||
| Read Oct 3 | Layer MCP, Blocking McpClient.executeTool(...), Non-blocking counterpart executeToolAsync(...) | McpClient has an executeToolAsync method | Holds |
McpClient has an executeTool method | Holds | ||
| Read Oct 3 | Layer Guardrail execution, Blocking ChatExecutor.execute(...), Non-blocking counterpart ChatExecutor.executeAsync(...) | ChatExecutor has an executeAsync method | Holds |
ChatExecutor has an execute method | Holds | ||
| Read Oct 3 | Layer HTTP client, Blocking execute(...), Non-blocking counterpart executeAsync(...) , stream(...) | HttpClient has an executeAsync method | Holds |
HttpClient has a stream method | Holds | ||
HttpClient has an execute method | Holds | ||
| Read Oct 3 | Everywhere LangChain4j runs work off the caller thread — concurrent tool calls, offloaded retrieval, retry backoff — it takes the executor from one pluggable seam, the ExecutorProvider SPI: | ExecutorProvider is a contract | Holds |
| Read Oct 3 | Provider OpenAI — Chat Completions, CompletableFuture ( chatAsync ) ✅, Flow.Publisher (reactive chat ) ✅ | OpenAiChatModel has chatAsync | Holds |
OpenAiStreamingChatModel has chat | Holds | ||
| Read Oct 3 | Provider OpenAI — Responses, CompletableFuture ( chatAsync ) ✅, Flow.Publisher (reactive chat ) ✅ | OpenAiOfficialChatModel has chatAsync | Holds |
OpenAiOfficialStreamingChatModel has chat | Holds | ||
| Read Oct 3 | Provider Anthropic, CompletableFuture ( chatAsync ) ✅, Flow.Publisher (reactive chat ) ✅ | AnthropicChatModel has chatAsync | Holds |
AnthropicStreamingChatModel has chat | Holds | ||
| Read Oct 3 | Provider Bedrock, CompletableFuture ( chatAsync ) ✅, Flow.Publisher (reactive chat ) ✅ | BedrockChatModel has chatAsync | Holds |
BedrockStreamingChatModel has chat | Holds | ||
| Read Oct 3 | OpenAI implements asynchronous embeddings ( embedAsync), Cohere asynchronous scoring (scoreAsync) and Tavily asynchronous web search (searchAsync). | OpenAiEmbeddingModel has embedAsync | Holds |
| Read Oct 3 | Component Chat memory, What it must implement addAsync / messagesAsync / setAsync, Bundled implementations MessageWindowChatMemory and InMemoryChatMemoryStore already do | MessageWindowChatMemory has addAsync | Holds |
MessageWindowChatMemory has messagesAsync | Holds | ||
MessageWindowChatMemory has setAsync | Holds | ||
InMemoryChatMemoryStore has messagesAsync | Holds | ||
| Read Oct 3 | Component Chat model router ( RoutingChatModel ), What it must implement routeAsync, Bundled implementations DecisionModelChatModelRouter does, if its decision model supports decideAsync | DecisionModelChatModelRouter has routeAsync | Holds |
| Read Oct 3 | Each type of event includes information encapsulated inside an InvocationContext. | AiServiceStartedEvent depends on InvocationContext | Holds |
AiServiceErrorEvent depends on InvocationContext | Holds | ||
AiServiceCompletedEvent depends on InvocationContext | Holds | ||
ToolExecutedEvent depends on InvocationContext | Holds | ||
| Read Oct 3 | Listener Name AiServiceStartedListener , Event AiServiceStartedEvent | AiServiceStartedListener depends on AiServiceStartedEvent | Holds |
| Read Oct 3 | Listener Name AiServiceErrorListener , Event AiServiceErrorEvent | AiServiceErrorListener depends on AiServiceErrorEvent | Holds |
| Read Oct 3 | Listener Name AiServiceCompletedListener , Event AiServiceCompletedEvent | AiServiceCompletedListener depends on AiServiceCompletedEvent | Holds |
| Read Oct 3 | Listener Name ToolExecutedEventListener , Event ToolExecutedEvent | ToolExecutedEventListener depends on ToolExecutedEvent | Holds |
| Read Oct 3 | There are various registerListener method variants on the AiServices class. | AiServices has a registerListener method | Holds |
| Read Oct 3 | Once you have your event and listener, you need to fire the event by obtaining/managing an instance of AiServiceListenerRegistrar and calling the fireEvent(event) method. | AiServiceListenerRegistrar has a fireEvent method | Holds |
| Read Oct 3 | The attributes map allows passing information between the onRequest, onResponse, and onError methods of the same ChatModelListener, as well as between multiple ChatModelListeners. | ChatModelListener has an onRequest method | Holds |
ChatModelListener has an onResponse method | Holds | ||
ChatModelListener has an onError method | Holds | ||
| Read Oct 3 | For example, you can pass tenant or correlation identifiers to ChatModelListeners through listenerAttributes. | ChatRequestOptions has listenerAttributes | Holds |
| Read Oct 3 | The EmbeddingStoreIngestor loads a DocumentSplitter from the langchain4j-easy-rag dependency through SPI. | EmbeddingStoreIngestor depends on DocumentSplitter | Holds |
| Read Oct 3 | The EmbeddingStoreIngestor loads an EmbeddingModel from the langchain4j-easy-rag dependency through SPI. | EmbeddingStoreIngestor depends on EmbeddingModel | Holds |
| Read Oct 3 | Documents can represent files in various formats, such as PDF, DOC, TXT, etc. To parse each of these formats, there's a DocumentParser interface with several implementations included in the library: | DocumentParser is a contract | Holds |
| Read Oct 3 | DocumentTransformer implementations can perform a variety of document transformations such as: | DocumentTransformer is a contract | Holds |
| Read Oct 3 | Currently, the only implementation provided out-of-the-box is HtmlToTextDocumentTransformer in the langchain4j-document-transformer-jsoup module, which can extract desired text content and metadata entries from the raw HTML. | HtmlToTextDocumentTransformer implements DocumentTransformer | Holds |
| Read Oct 3 | TextSegment.text() returns the text of the TextSegment | TextSegment has a text method | Holds |
| Read Oct 3 | TextSegment.metadata() returns the Metadata of the TextSegment | TextSegment has a metadata method | Holds |
| Read Oct 3 | TextSegment.from(String, Metadata) creates a TextSegment from text and Metadata | TextSegment has a from method | Holds |
| Read Oct 3 | LangChain4j has a DocumentSplitter interface with several out-of-the-box implementations: | DocumentSplitter is a contract | Holds |
| Read Oct 3 | You call the split(Document) or splitAll(List<Document>) methods of the DocumentSplitter. | DocumentSplitter has a split method | Holds |
DocumentSplitter has a splitAll method | Holds | ||
| Read Oct 3 | For instance, DocumentByParagraphSplitter divides a document into paragraphs (defined by two or more consecutive newline characters), while DocumentBySentenceSplitter uses the OpenNLP library's sentence detector to split a document into sentences, and so on. | DocumentByParagraphSplitter implements DocumentSplitter | Holds |
DocumentBySentenceSplitter implements DocumentSplitter | Holds | ||
| Read Oct 3 | CosineSimilarity.between(Embedding, Embedding) calculates the cosine similarity between 2 Embeddings | CosineSimilarity has a between method | Holds |
| Read Oct 3 | The EmbeddingModel interface represents a special type of model that converts text into an Embedding. | EmbeddingModel is a contract | Holds |
| Read Oct 3 | EmbeddingModel.embed(String) embeds the given text | EmbeddingModel has an embed method | Holds |
| Read Oct 3 | EmbeddingModel.embedAll(List<TextSegment>) embeds all the given TextSegment | EmbeddingModel has an embedAll method | Holds |
| Read Oct 3 | EmbeddingModel.dimension() returns the dimension of the Embedding produced by this model </details> | EmbeddingModel has a dimension method | Holds |
| Read Oct 3 | All ingest() methods in EmbeddingStoreIngestor return an IngestionResult. | EmbeddingStoreIngestor has an ingest method | Holds |
| Read Oct 3 | DefaultRetrievalAugmentor, which should be suitable for the majority of RAG use cases. | DefaultRetrievalAugmentor implements RetrievalAugmentor | Holds |
| Read Oct 3 | EmbeddingStoreContentRetriever retrieves relevant Content from the EmbeddingStore using the EmbeddingModel to embed the Query. | EmbeddingStoreContentRetriever depends on EmbeddingStore | Holds |
EmbeddingStoreContentRetriever depends on EmbeddingModel | Holds | ||
| Read Oct 3 | WebSearchContentRetriever retrieves relevant Content from the web using a WebSearchEngine. | WebSearchContentRetriever depends on WebSearchEngine | Holds |
| Read Oct 3 | SqlDatabaseContentRetriever is an experimental implementation of the ContentRetriever that can be found in the langchain4j-experimental-sql module. | SqlDatabaseContentRetriever implements ContentRetriever | Holds |
| Read Oct 3 | DefaultQueryRouter is the default implementation used in DefaultRetrievalAugmentor. | DefaultQueryRouter implements QueryRouter | Holds |
DefaultRetrievalAugmentor depends on DefaultQueryRouter | Holds | ||
| Read Oct 3 | The DefaultContentAggregator is the default implementation of ContentAggregator, which performs two-stage Reciprocal Rank Fusion (RRF). | DefaultContentAggregator implements ContentAggregator | Holds |
| Read Oct 3 | The ReRankingContentAggregator uses a ScoringModel, like Cohere, to perform re-ranking. | ReRankingContentAggregator depends on ScoringModel | Holds |
| Read Oct 3 | There are several ways to add properties to a JsonObjectSchema: You can add all the properties at once using the properties(Map<String, JsonSchemaElement> properties) method: | JsonObjectSchema implements JsonSchemaElement | Holds |
| Read Oct 3 | The structure of the JSON schema is defined using JsonSchemaElement interface, with the following subtypes: | JsonSchemaElement is a contract | Holds |
| Read Oct 3 | You can add all the properties at once using the properties(Map<String, JsonSchemaElement> properties) method: | JsonObjectSchema has a properties method | Holds |
| Read Oct 3 | All of the JsonSchemaElement subtypes, except for JsonReferenceSchema, have a description property. | JsonAnyOfSchema has description | Holds |
JsonArraySchema has description | Holds | ||
JsonBooleanSchema has description | Holds | ||
JsonEnumSchema has description | Holds | ||
JsonIntegerSchema has description | Holds | ||
JsonNullSchema has description | Holds | ||
JsonNumberSchema has description | Holds | ||
JsonObjectSchema has description | Holds | ||
JsonRawSchema has description | Holds | ||
JsonStringSchema has description | Holds | ||
| Read Oct 3 | At the low level, you can use the chat(ChatRequest) method of the ChatModel. | ChatModel has a chat method | Holds |
ChatModel depends on ChatRequest | Holds | ||
| Read Oct 3 | At the low level, you can use the chat(ChatRequest) method of the ChatModel. A similar method is also present in the StreamingChatModel. | StreamingChatModel has a chat method | Holds |
| Read Oct 3 | The description of the tool | ToolSpecification has description | Holds |
| Read Oct 3 | The parameters of the tool and their descriptions | ToolSpecification has parameters | Holds |
| Read Oct 3 | The metadata of the tool. | ToolSpecification has metadata | Holds |
| Read Oct 3 | ToolSpecifications.toolSpecificationsFrom(Class) ToolSpecifications.toolSpecificationsFrom(Object) | ToolSpecifications has a toolSpecificationsFrom method | Holds |
| Read Oct 3 | ToolSpecifications.toolSpecificationFrom(Method) | ToolSpecifications has a toolSpecificationFrom method | Holds |
| Read Oct 3 | ToolSpecification can be serialized to JSON and deserialized back using toJson() and fromJson() methods. | ToolSpecification has a toJson method | Holds |
ToolSpecification has a fromJson method | Holds | ||
| Read Oct 3 | You can provide your own implementation of ToolSpecificationJsonCodec via SPI by implementing ToolSpecificationJsonCodecFactory and registering it in META-INF/services/dev.langchain4j.spi.agent.tool.ToolSpecificationJsonCodecFactory. | ToolSpecificationJsonCodec is a contract | Holds |
ToolSpecificationJsonCodecFactory is a contract | Holds | ||
| Read Oct 3 | If the LLM decides to call the tool, the returned AiMessage will contain data in the toolExecutionRequests field. | AiMessage has a toolExecutionRequests field | Holds |
| Read Oct 3 | In this case, AiMessage.hasToolExecutionRequests() will return true. Depending on the LLM, it can contain one or multiple ToolExecutionRequest objects (some LLMs support calling multiple tools in parallel). | AiMessage depends on ToolExecutionRequest | Holds |
| Read Oct 3 | In this case, AiMessage.hasToolExecutionRequests() will return true. | AiMessage has a hasToolExecutionRequests method | Holds |
| Read Oct 3 | The name of the tool to be called, for example: | ToolExecutionRequest has name | Holds |
| Read Oct 3 | The arguments, for example: { "city": "London", "temperatureUnit": "CELSIUS" } | ToolExecutionRequest has arguments | Holds |
| Read Oct 3 | Instead of using text(), you can use the builder with contents(): | ToolExecutionResultMessage has a text method | Holds |
| Read Oct 3 | If the LLM initiates multiple tool calls, the index will increment, allowing you to correlate the different PartialToolCalls with each other and with the final CompleteToolCall. | PartialToolCall has index | Holds |
CompleteToolCall has index | Holds | ||
| Read Oct 3 | The ToolProvider accepts a ToolProviderRequest (that contains the UserMessage, chat memory ID and InvocationParameters) and returns a ToolProviderResult that contains tools for the current AI Service invocation. | ToolProvider depends on ToolProviderRequest | Holds |
ToolProvider depends on ToolProviderResult | Holds | ||
| Read Oct 3 | When building ToolProviderResult, you can mark tools for immediate return using ToolProviderResult.builder(). | ToolProviderResult has a builder method | Holds |
| Read Oct 3 | Tool search is implemented via the ToolSearchStrategy interface: | ToolSearchStrategy is a contract | Holds |
| Read Oct 3 | SimpleToolSearchStrategy – keyword-based matching | SimpleToolSearchStrategy implements ToolSearchStrategy | Holds |
| Read Oct 3 | VectorToolSearchStrategy – semantic search using embeddings | VectorToolSearchStrategy implements ToolSearchStrategy | Holds |
| Read Oct 3 | DecisionModelToolSearchStrategy – a decision model decides which tools match the search | DecisionModelToolSearchStrategy implements ToolSearchStrategy | Holds |
| Read Oct 3 | When using MCP tools (via McpToolProvider), always-visible tools can be configured via alwaysVisibleToolNames: | McpToolProvider has alwaysVisibleToolNames | Holds |
| Read Oct 3 | This can be done by configuring the returnBehavior field of the @Tool annotation as in the following example: | Tool has a returnBehavior field | Holds |
| Read Oct 3 | ReturnBehavior.IMMEDIATE_IF_LAST is meant for tools the LLM uses to explicitly signal the end of a multi-step action — for example, an endExecutionAndGetFinalResult tool the LLM appendsafter a sequence of clicks, navigations, etc. | ReturnBehavior has IMMEDIATE_IF_LAST | Holds |
| Read Oct 3 | a response made up only of IMMEDIATE and/or IMMEDIATE_IF_LAST tools returns immediately regardless of which one is last (still subject to the no-errors rule). | ReturnBehavior has IMMEDIATE | Holds |
| Read Oct 3 | Handler ToolArgumentsErrorHandler.sendExceptionMessageToLlm(), What happens when a tool fails The message of the error is sent to the LLM, so that it can correct the arguments and try again The AI Service invocation continues ⚠️ See the warning below | ToolArgumentsErrorHandler has a sendExceptionMessageToLlm method | Holds |
| Read Oct 3 | Handler ToolArgumentsErrorHandler.failInvocation(), What happens when a tool fails The error is rethrown the AI Service invocation fails and nothing is sent to the LLM | ToolArgumentsErrorHandler has a failInvocation method | Holds |
| Read Oct 3 | Handler ToolExecutionErrorHandler.sendExceptionMessageToLlm(), What happens when a tool fails The message of the exception thrown by the tool is sent to the LLM (or the text from messageForLlm() , if the exception implements ToolErrorVisibleToLlm ), so that it can react to it The AI Service invocation continues ⚠️ See the warning below | ToolExecutionErrorHandler has a sendExceptionMessageToLlm method | Holds |
| Read Oct 3 | Handler ToolExecutionErrorHandler.failInvocationUnlessVisibleToLlm(), What happens when a tool fails Only exceptions implementing ToolErrorVisibleToLlm are shown to the LLM, using the text they provide Every other exception fails the AI Service invocation See Deciding per exception what the LLM sees | ToolExecutionErrorHandler has a failInvocationUnlessVisibleToLlm method | Holds |
| Read Oct 3 | Handler ToolExecutionErrorHandler.sendGenericMessageToLlmUnlessVisibleToLlm(String), What happens when a tool fails Exceptions implementing ToolErrorVisibleToLlm are shown to the LLM, using the text they provide For every other exception, the given generic message is sent to the LLM, and the exception is logged at WARN level The AI Service invocation continues | ToolExecutionErrorHandler has a sendGenericMessageToLlmUnlessVisibleToLlm method | Holds |
| Read Oct 3 | Handler ToolExecutionErrorHandler.failInvocation(), What happens when a tool fails The exception is rethrown the AI Service invocation fails and nothing is sent to the LLM | ToolExecutionErrorHandler has a failInvocation method | Holds |
| Read Oct 3 | Each ToolExecutionRequest should contain: The id of the tool call. | ToolExecutionRequest has id | Holds |
| Read Oct 3 | The tools that LangChain4j provides, such as skills, tool search and MCP tools, report the errors the LLM can do something about (for example, a missing argument or an unknown skill name) with LlmVisibleToolExecutionException, which implements ToolErrorVisibleToLlm, so the LLM keeps seeing them with this handler too. | LlmVisibleToolExecutionException implements ToolErrorVisibleToLlm | Holds |
345 rules from 17 docs
See these where the change is. The browser extension puts a pull request's doc findings, and a diagram of what it changed, on the GitHub page itself, so the code and what your docs say about it are side by side.