Cohere 통합

커뮤니티에서 만든 Cohere 채팅 모델 통합 문서예요. Cohere의 V2 Chat API를 기반으로 구현됐어요. Cohere의 command-r 계열 모델을 LangChain4j에서 바로 쓸 수 있는데, 일반 채팅은 CohereChatModel, 스트리밍은 CohereStreamingChatModel을 주로 쓰게 돼요.

출처: 공식문서

Maven 의존성

1.0.0-alpha1 이후:

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-community-cohere</artifactId>
    <version>${latest version here}</version>
</dependency>

BOM을 쓰면 의존성 버전을 일관되게 관리할 수도 있어요:

<dependencyManagement>
    <dependency>
        <groupId>dev.langchain4j</groupId>
        <artifactId>langchain4j-community-bom</artifactId>
        <version>${latest version here}</version>
        <type>pom</type>
        <scope>import</scope>
    </dependency>
</dependencyManagement>

채팅 모델 지원

CohereChatModel은 다음 코드로 만들 수 있어요:

ChatModel model = CohereChatModel.builder()
        .apiKey(System.getenv("CO_API_KEY"))
        .modelName("command-r7b-12-2024")
        .logRequests(true)
        .logResponses(true)
        .build();

스트리밍 응답에는 CohereStreamingChatModel을 써요:

StreamingChatModel streamingModel = CohereStreamingChatModel.builder()
        .apiKey(System.getenv("CO_API_KEY"))
        .modelName("command-r7b-12-2024")
        .logRequests(true)
        .logResponses(true)
        .build();

설정 가능한 파라미터

CohereChatModel과 CohereStreamingChatModel은 다음 파라미터를 받아요.

Property	Description	Default Value
`baseUrl`	The URL to connect to the Cohere API.	https://api.cohere.com/v2/
`apiKey`	The API Key.
`modelName`	The model to use, e.g. `command-r7b-12-2024` or `command-r-plus`.
`timeout`	HTTP client timeout for requests.
`maxRetries`	Maximum number of retries per request. Only available on `CohereChatModel`.	3
`temperature`	Sampling temperature.
`topP`	Nucleus sampling threshold.
`topK`	Limits sampling to the `topK` most likely tokens at each step.
`frequencyPenalty`	Penalty for tokens based on how often they have appeared.
`presencePenalty`	Penalty for tokens that have appeared at least once.
`maxTokens`	The maximum number of tokens returned by this request.
`stopSequences`	Sequences that cause the model to stop generating further text.
`toolSpecifications`	Tool (function) definitions the model can call.
`toolChoice`	A `ToolChoice` controlling how the model selects tools. Possible values: `AUTO`, `REQUIRED`.
`responseFormat`	The response format, e.g. `TEXT` or `JSON`.
`thinkingType`	A `CohereThinkingType` enabling or disabling extended thinking for reasoning-capable models.
`thinkingTokenBudget`	Maximum tokens the model may spend on internal thinking.
`safetyMode`	A `CohereSafetyMode` inserted into the prompt. Possible values: `CONTEXTUAL`, `STRICT`, `OFF`.
`priority`	Request priority when the Cohere API is under load.
`seed`	If set, the model samples tokens deterministically.
`logprobs`	Whether to include token log probabilities in the response.
`strictTools`	Whether to enforce strict adherence to tool definitions.
`defaultRequestParameters`	Default `ChatRequestParameters` applied to every request.
`listeners`	Listeners that listen for request, response and errors.
`logRequests`	Whether to log request or not.	`false`
`logResponses`	Whether to log response or not.	`false`

표의 값은 영문 그대로 보존했어요. 표를 보면 maxRetries는 CohereChatModel에서만 쓸 수 있고 기본이 3, baseUrl 기본값은 https://api.cohere.com/v2/라는 걸 알 수 있어요.

응답 메타데이터

Cohere 전용 응답 메타데이터에 접근할 수 있어요:

ChatResponse response = model.chat(UserMessage.from("Hello"));
CohereChatResponseMetadata metadata = (CohereChatResponseMetadata) response.metadata();

List<CohereLogprobs> logprobs = metadata.logprobs();
CohereBilledUnits billedUnits = metadata.billedUnits();
Integer cachedTokens = metadata.cachedTokens();

Property	Description
`logprobs`	Log probabilities for generated tokens. Returned when `logprobs` is enabled.
`billedUnits`	Billing breakdown for the request (input tokens, output tokens, search units, classifications).
`cachedTokens`	Number of tokens served from Cohere's prompt cache.

logprobs는 logprobs 옵션을 켰을 때만 돌아오고, billedUnits는 요청의 과금 내역(입력·출력 토큰, 검색 단위, 분류), cachedTokens는 Cohere 프롬프트 캐시에서 제공된 토큰 수예요.

Cohere 통합

Cohere 통합

Maven 의존성

채팅 모델 지원

설정 가능한 파라미터

응답 메타데이터

예시

더 알아보기