Merge branch 'main' into build-platform-matrix

core/build.gradle.kts

@@ -68,8 +68,6 @@ kotlin {
         api(libs.bundles.ktor.client)
         api(projects.xefTokenizer)
 
-        // implementation(libs.arrow.fx.stm)
-
         implementation(libs.uuid)
         implementation(libs.klogging)
       }

core/src/commonMain/kotlin/com/xebia/functional/xef/AIError.kt

@@ -1,7 +1,7 @@
 package com.xebia.functional.xef
 
 import arrow.core.NonEmptyList
-import com.xebia.functional.xef.llm.openai.Message
+import com.xebia.functional.xef.llm.models.chat.Message
 import kotlin.jvm.JvmOverloads
 
 sealed class AIError @JvmOverloads constructor(message: String, cause: Throwable? = null) :

core/src/commonMain/kotlin/com/xebia/functional/xef/auto/AI.kt

@@ -1,18 +1,8 @@
 package com.xebia.functional.xef.auto
 
-import arrow.core.Either
-import arrow.core.left
-import arrow.core.right
 import com.xebia.functional.xef.AIError
-import com.xebia.functional.xef.embeddings.OpenAIEmbeddings
-import com.xebia.functional.xef.env.OpenAIConfig
-import com.xebia.functional.xef.llm.openai.AIClient
-import com.xebia.functional.xef.llm.openai.LLMModel
-import com.xebia.functional.xef.llm.openai.MockOpenAIClient
-import com.xebia.functional.xef.llm.openai.simpleMockAIClient
-import com.xebia.functional.xef.vectorstores.LocalVectorStore
+import com.xebia.functional.xef.llm.AIClient
 import com.xebia.functional.xef.vectorstores.VectorStore
-import kotlin.time.ExperimentalTime
 
 @DslMarker annotation class AiDsl
 
@@ -29,79 +19,9 @@ typealias AI<A> = suspend CoreAIScope.() -> A
 /** A DSL block that makes it more convenient to construct [AI] values. */
 inline fun <A> ai(noinline block: suspend CoreAIScope.() -> A): AI<A> = block
 
-/**
- * Run the [AI] value to produce an [A], this method initialises all the dependencies required to
- * run the [AI] value and once it finishes it closes all the resources.
- *
- * This operator is **terminal** meaning it runs and completes the _chain_ of `AI` actions.
- */
-suspend inline fun <A> AI<A>.getOrElse(
-  runtime: AIRuntime<A> = AIRuntime.openAI(),
-  crossinline orElse: suspend (AIError) -> A
-): A = AIScope(runtime, this) { orElse(it) }
-
 suspend fun <A> AIScope(runtime: AIRuntime<A>, block: AI<A>, orElse: suspend (AIError) -> A): A =
   try {
     runtime.runtime(block)
   } catch (e: AIError) {
     orElse(e)
   }
-
-@OptIn(ExperimentalTime::class)
-suspend fun <A> MockAIScope(
-  mockClient: MockOpenAIClient,
-  block: suspend CoreAIScope.() -> A,
-  orElse: suspend (AIError) -> A
-): A =
-  try {
-    val embeddings = OpenAIEmbeddings(OpenAIConfig(), mockClient)
-    val vectorStore = LocalVectorStore(embeddings)
-    val scope =
-      CoreAIScope(
-        LLMModel.GPT_3_5_TURBO,
-        LLMModel.GPT_3_5_TURBO_FUNCTIONS,
-        mockClient,
-        vectorStore,
-        embeddings
-      )
-    block(scope)
-  } catch (e: AIError) {
-    orElse(e)
-  }
-
-/**
- * Run the [AI] value to produce _either_ an [AIError], or [A]. this method initialises all the
- * dependencies required to run the [AI] value and once it finishes it closes all the resources.
- *
- * This operator is **terminal** meaning it runs and completes the _chain_ of `AI` actions.
- *
- * @see getOrElse for an operator that allow directly handling the [AIError] case.
- */
-suspend inline fun <reified A> AI<A>.toEither(): Either<AIError, A> =
-  ai { invoke().right() }.getOrElse { it.left() }
-
-/**
- * Run the [AI] value to produce _either_ an [AIError], or [A]. This method uses the [mockAI] to
- * compute the different responses.
- */
-suspend fun <A> AI<A>.mock(mockAI: MockOpenAIClient): Either<AIError, A> =
-  MockAIScope(mockAI, { invoke().right() }, { it.left() })
-
-/**
- * Run the [AI] value to produce _either_ an [AIError], or [A]. This method uses the [mockAI] to
- * compute the different responses.
- */
-suspend fun <A> AI<A>.mock(mockAI: (String) -> String): Either<AIError, A> =
-  MockAIScope(simpleMockAIClient(mockAI), { invoke().right() }, { it.left() })
-
-/**
- * Run the [AI] value to produce [A]. this method initialises all the dependencies required to run
- * the [AI] value and once it finishes it closes all the resources.
- *
- * This operator is **terminal** meaning it runs and completes the _chain_ of `AI` actions.
- *
- * @throws AIError in case something went wrong.
- * @see getOrElse for an operator that allow directly handling the [AIError] case instead of
- *   throwing.
- */
-suspend inline fun <reified A> AI<A>.getOrThrow(): A = getOrElse { throw it }

core/src/commonMain/kotlin/com/xebia/functional/xef/auto/AIRuntime.kt

@@ -1,32 +1,12 @@
 package com.xebia.functional.xef.auto
 
-import com.xebia.functional.xef.embeddings.OpenAIEmbeddings
-import com.xebia.functional.xef.env.OpenAIConfig
-import com.xebia.functional.xef.llm.openai.KtorOpenAIClient
-import com.xebia.functional.xef.llm.openai.LLMModel
-import com.xebia.functional.xef.vectorstores.LocalVectorStore
-import kotlin.jvm.JvmStatic
-import kotlin.time.ExperimentalTime
+import com.xebia.functional.xef.embeddings.Embeddings
+import com.xebia.functional.xef.llm.AIClient
 
-data class AIRuntime<A>(val runtime: suspend (block: AI<A>) -> A) {
-  companion object {
-    @OptIn(ExperimentalTime::class)
-    @JvmStatic
-    fun <A> openAI(): AIRuntime<A> = AIRuntime { block ->
-      val openAIConfig = OpenAIConfig()
-      KtorOpenAIClient(openAIConfig).use { openAiClient ->
-        val embeddings = OpenAIEmbeddings(openAIConfig, openAiClient)
-        val vectorStore = LocalVectorStore(embeddings)
-        val scope =
-          CoreAIScope(
-            defaultModel = LLMModel.GPT_3_5_TURBO_16K,
-            defaultSerializationModel = LLMModel.GPT_3_5_TURBO_FUNCTIONS,
-            aiClient = openAiClient,
-            context = vectorStore,
-            embeddings = embeddings
-          )
-        block(scope)
-      }
-    }
-  }
+data class AIRuntime<A>(
+  val client: AIClient,
+  val embeddings: Embeddings,
+  val runtime: suspend (block: AI<A>) -> A
+) {
+  companion object
 }

core/src/commonMain/kotlin/com/xebia/functional/xef/auto/CoreAIScope.kt

@@ -7,10 +7,17 @@ import com.xebia.functional.tokenizer.ModelType
 import com.xebia.functional.tokenizer.truncateText
 import com.xebia.functional.xef.AIError
 import com.xebia.functional.xef.embeddings.Embeddings
-import com.xebia.functional.xef.llm.openai.*
-import com.xebia.functional.xef.llm.openai.CFunction
-import com.xebia.functional.xef.llm.openai.images.ImagesGenerationRequest
-import com.xebia.functional.xef.llm.openai.images.ImagesGenerationResponse
+import com.xebia.functional.xef.llm.AIClient
+import com.xebia.functional.xef.llm.LLM
+import com.xebia.functional.xef.llm.LLMModel
+import com.xebia.functional.xef.llm.models.chat.ChatCompletionRequest
+import com.xebia.functional.xef.llm.models.chat.ChatCompletionRequestWithFunctions
+import com.xebia.functional.xef.llm.models.chat.Message
+import com.xebia.functional.xef.llm.models.chat.Role
+import com.xebia.functional.xef.llm.models.functions.CFunction
+import com.xebia.functional.xef.llm.models.images.ImagesGenerationRequest
+import com.xebia.functional.xef.llm.models.images.ImagesGenerationResponse
+import com.xebia.functional.xef.llm.models.text.CompletionRequest
 import com.xebia.functional.xef.prompt.Prompt
 import com.xebia.functional.xef.vectorstores.CombinedVectorStore
 import com.xebia.functional.xef.vectorstores.LocalVectorStore
@@ -25,8 +32,8 @@ import kotlin.jvm.JvmName
  * programs.
  */
 class CoreAIScope(
-  val defaultModel: LLMModel,
-  val defaultSerializationModel: LLMModel,
+  val defaultModel: LLM.Chat,
+  val defaultSerializationModel: LLM.ChatWithFunctions,
   val aiClient: AIClient,
   val context: VectorStore,
   val embeddings: Embeddings,
@@ -117,7 +124,7 @@ class CoreAIScope(
     functions: List<CFunction>,
     serializer: (json: String) -> A,
     maxDeserializationAttempts: Int = this.maxDeserializationAttempts,
-    model: LLMModel = defaultSerializationModel,
+    model: LLM.ChatWithFunctions = defaultSerializationModel,
     user: String = this.user,
     echo: Boolean = this.echo,
     numberOfPredictions: Int = this.numberOfPredictions,
@@ -162,7 +169,7 @@ class CoreAIScope(
   @AiDsl
   suspend fun promptMessage(
     question: String,
-    model: LLMModel = defaultModel,
+    model: LLM.Chat = defaultModel,
     functions: List<CFunction> = emptyList(),
     user: String = this.user,
     echo: Boolean = this.echo,
@@ -186,7 +193,7 @@ class CoreAIScope(
   @AiDsl
   suspend fun promptMessage(
     prompt: Prompt,
-    model: LLMModel = defaultModel,
+    model: LLM.Chat = defaultModel,
     functions: List<CFunction> = emptyList(),
     user: String = this.user,
     echo: Boolean = this.echo,
@@ -249,7 +256,7 @@ class CoreAIScope(
     }
 
     suspend fun buildChatRequest(): ChatCompletionRequest {
-      val messages: List<Message> = listOf(Message(Role.system.name, promptWithContext))
+      val messages: List<Message> = listOf(Message(Role.SYSTEM.name, promptWithContext))
       return ChatCompletionRequest(
         model = model.name,
         user = user,
@@ -261,7 +268,7 @@ class CoreAIScope(
     }
 
     suspend fun chatWithFunctionsRequest(): ChatCompletionRequestWithFunctions {
-      val role: String = Role.user.name
+      val role: String = Role.USER.name
       val firstFnName: String? = functions.firstOrNull()?.name
       val messages: List<Message> = listOf(Message(role, promptWithContext))
       return ChatCompletionRequestWithFunctions(
@@ -276,15 +283,15 @@ class CoreAIScope(
       )
     }
 
-    return when (model.kind) {
-      LLMModel.Kind.Completion ->
+    return when (model) {
+      is LLM.Completion ->
         aiClient.createCompletion(buildCompletionRequest()).choices.map { it.text }
-      LLMModel.Kind.Chat ->
-        aiClient.createChatCompletion(buildChatRequest()).choices.map { it.message.content }
-      LLMModel.Kind.ChatWithFunctions ->
-        aiClient.createChatCompletionWithFunctions(chatWithFunctionsRequest()).choices.map {
-          it.message.functionCall.arguments
+      is LLM.ChatWithFunctions ->
+        aiClient.createChatCompletionWithFunctions(chatWithFunctionsRequest()).choices.mapNotNull {
+          it.message?.functionCall?.arguments
         }
+      else ->
+        aiClient.createChatCompletion(buildChatRequest()).choices.mapNotNull { it.message?.content }
     }
   }
 
@@ -323,16 +330,16 @@ class CoreAIScope(
     } else prompt
   }
 
-  private fun tokensFromMessages(messages: List<Message>, model: LLMModel): Int {
+  private fun tokensFromMessages(messages: List<Message>, model: LLM): Int {
     fun Encoding.countTokensFromMessages(tokensPerMessage: Int, tokensPerName: Int): Int =
       messages.sumOf { message ->
         countTokens(message.role) +
-          countTokens(message.content) +
+          (message.content?.let { countTokens(it) } ?: 0) +
           tokensPerMessage +
           (message.name?.let { tokensPerName } ?: 0)
       } + 3
 
-    fun fallBackTo(fallbackModel: LLMModel, paddingTokens: Int): Int {
+    fun fallBackTo(fallbackModel: LLM, paddingTokens: Int): Int {
       logger.debug {
         "Warning: ${model.name} may change over time. " +
           "Returning messages num tokens assuming ${fallbackModel.name} + $paddingTokens padding tokens."

core/src/commonMain/kotlin/com/xebia/functional/xef/embeddings/Embeddings.kt

@@ -1,6 +1,6 @@
 package com.xebia.functional.xef.embeddings
 
-import com.xebia.functional.xef.llm.openai.RequestConfig
+import com.xebia.functional.xef.llm.models.embeddings.RequestConfig
 
 data class Embedding(val data: List<Float>)
 

core/src/commonMain/kotlin/com/xebia/functional/xef/llm/AIClient.kt

@@ -0,0 +1,28 @@
+package com.xebia.functional.xef.llm
+
+import com.xebia.functional.xef.llm.models.chat.ChatCompletionRequest
+import com.xebia.functional.xef.llm.models.chat.ChatCompletionRequestWithFunctions
+import com.xebia.functional.xef.llm.models.chat.ChatCompletionResponse
+import com.xebia.functional.xef.llm.models.chat.ChatCompletionResponseWithFunctions
+import com.xebia.functional.xef.llm.models.embeddings.EmbeddingRequest
+import com.xebia.functional.xef.llm.models.embeddings.EmbeddingResult
+import com.xebia.functional.xef.llm.models.images.ImagesGenerationRequest
+import com.xebia.functional.xef.llm.models.images.ImagesGenerationResponse
+import com.xebia.functional.xef.llm.models.text.CompletionRequest
+import com.xebia.functional.xef.llm.models.text.CompletionResult
+
+interface AIClient : AutoCloseable {
+  suspend fun createCompletion(request: CompletionRequest): CompletionResult
+
+  suspend fun createChatCompletion(request: ChatCompletionRequest): ChatCompletionResponse
+
+  suspend fun createChatCompletionWithFunctions(
+    request: ChatCompletionRequestWithFunctions
+  ): ChatCompletionResponseWithFunctions
+
+  suspend fun createEmbeddings(request: EmbeddingRequest): EmbeddingResult
+
+  suspend fun createImages(request: ImagesGenerationRequest): ImagesGenerationResponse
+
+  override fun close() {}
+}