query fine tuned models (#469)

* query fine tuned models (from branch #441-fine-tuning-oai)

* spotless

* clean build, make tests and mocks compile

* changes according to pr comments

---------

Co-authored-by: José Carlos Montañez <josecarlos.montanez@xebia.com>

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

@@ -12,6 +12,12 @@ sealed interface LLM : AutoCloseable {
   val name
     get() = modelType.name
 
+  /**
+   * Copies this instance and uses [modelType] for [LLM.modelType]. Has to return the most specific
+   * type of this instance!
+   */
+  fun copy(modelType: ModelType): LLM
+
   fun tokensFromMessages(
     messages: List<Message>
   ): Int { // TODO: naive implementation with magic numbers

core/src/commonTest/kotlin/com/xebia/functional/xef/data/TestEmbeddings.kt

@@ -12,6 +12,8 @@ class TestEmbeddings : Embeddings {
 
   override val modelType: ModelType = ModelType.TODO("test-embeddings")
 
+  override fun copy(modelType: ModelType) = TestEmbeddings()
+
   override suspend fun embedDocuments(
     texts: List<String>,
     requestConfig: RequestConfig,

core/src/commonTest/kotlin/com/xebia/functional/xef/data/TestFunctionsModel.kt

@@ -18,6 +18,8 @@ class TestFunctionsModel(
 
   var requests: MutableList<FunChatCompletionRequest> = mutableListOf()
 
+  override fun copy(modelType: ModelType) = TestFunctionsModel(modelType, responses)
+
   override fun tokensFromMessages(messages: List<Message>): Int {
     return messages.sumOf { it.content.length }
   }

core/src/commonTest/kotlin/com/xebia/functional/xef/data/TestModel.kt

@@ -16,6 +16,8 @@ class TestModel(
 
   var requests: MutableList<ChatCompletionRequest> = mutableListOf()
 
+  override fun copy(modelType: ModelType) = TestModel(modelType, responses)
+
   override suspend fun createChatCompletion(
     request: ChatCompletionRequest
   ): ChatCompletionResponse {

examples/kotlin/src/main/kotlin/com/xebia/functional/xef/conversation/finetuning/FineTunedModelChat.kt

@@ -0,0 +1,23 @@
+package com.xebia.functional.xef.conversation.finetuning
+
+import com.xebia.functional.xef.conversation.llm.openai.OpenAI
+import com.xebia.functional.xef.env.getenv
+import com.xebia.functional.xef.prompt.Prompt
+
+suspend fun main() {
+  val spawnModelId =
+    getenv("OPENAI_FINE_TUNED_MODEL_ID")
+      ?: error("Please set the OPENAI_FINE_TUNED_MODEL_ID environment variable.")
+
+  val OAI = OpenAI()
+  val model = OAI.spawnModel(spawnModelId, OAI.GPT_3_5_TURBO)
+  OpenAI.conversation {
+    while (true) {
+      print("> ")
+      val question = readlnOrNull() ?: break
+      val answer = model.promptStreaming(Prompt(question), this)
+      answer.collect(::print)
+      println()
+    }
+  }
+}

gpt4all-kotlin/src/jvmMain/kotlin/com/xebia/functional/gpt4all/GPT4All.kt

@@ -66,6 +66,9 @@ interface GPT4All : AutoCloseable, Chat, Completion {
 
       val llModel = LLModel(path)
 
+      override fun copy(modelType: ModelType) =
+        GPT4All(url, path)
+
       override suspend fun createCompletion(request: CompletionRequest): CompletionResult =
         with(request) {
           val config = LLModel.config()

gpt4all-kotlin/src/jvmMain/kotlin/com/xebia/functional/gpt4all/HuggingFaceLocalEmbeddings.kt

@@ -11,13 +11,16 @@ import com.xebia.functional.xef.llm.models.usage.Usage
 
 class HuggingFaceLocalEmbeddings(
   override val modelType: ModelType,
-  artifact: String,
+  private val artifact: String,
 ) : Embeddings {
 
   private val tokenizer = HuggingFaceTokenizer.newInstance("${modelType.name}/$artifact")
 
   override val name: String = HuggingFaceLocalEmbeddings::class.java.canonicalName
 
+  override fun copy(modelType: ModelType) =
+    HuggingFaceLocalEmbeddings(modelType, artifact)
+
   override suspend fun createEmbeddings(request: EmbeddingRequest): EmbeddingResult {
     val embedings = tokenizer.batchEncode(request.input)
     return EmbeddingResult(

integrations/gcp/src/commonMain/kotlin/com/xebia/functional/xef/gcp/GCP.kt

@@ -47,10 +47,8 @@ class GCP(projectId: String? = null, location: VertexAIRegion? = null, token: St
 
   val defaultClient = GcpClient(config)
 
-  val CODECHAT by lazy { GcpChat(ModelType.TODO("codechat-bison@001"), defaultClient) }
-  val TEXT_EMBEDDING_GECKO by lazy {
-    GcpEmbeddings(ModelType.TODO("textembedding-gecko"), defaultClient)
-  }
+  val CODECHAT by lazy { GcpChat(this, ModelType.TODO("codechat-bison@001")) }
+  val TEXT_EMBEDDING_GECKO by lazy { GcpEmbeddings(this, ModelType.TODO("textembedding-gecko")) }
 
   @JvmField val DEFAULT_CHAT = CODECHAT
   @JvmField val DEFAULT_EMBEDDING = TEXT_EMBEDDING_GECKO

integrations/gcp/src/commonMain/kotlin/com/xebia/functional/xef/gcp/GcpClient.kt

@@ -79,7 +79,7 @@ class GcpClient(
       )
     val response =
       http.post(
-        "https://${config.location.officialName}-aiplatform.googleapis.com/v1/projects/${config.projectId}/locations/us-central1/publishers/google/models/$modelId:predict"
+        "https://${config.location.officialName}-aiplatform.googleapis.com/v1/projects/${config.projectId}/locations/${config.location.officialName}/publishers/google/models/$modelId:predict"
       ) {
         header("Authorization", "Bearer ${config.token}")
         contentType(ContentType.Application.Json)

integrations/gcp/src/commonMain/kotlin/com/xebia/functional/xef/gcp/models/GcpChat.kt

@@ -1,7 +1,7 @@
 package com.xebia.functional.xef.gcp.models
 
 import com.xebia.functional.tokenizer.ModelType
-import com.xebia.functional.xef.gcp.GcpClient
+import com.xebia.functional.xef.gcp.GCP
 import com.xebia.functional.xef.llm.Chat
 import com.xebia.functional.xef.llm.models.chat.*
 import com.xebia.functional.xef.llm.models.usage.Usage
@@ -13,10 +13,14 @@ import kotlinx.uuid.UUID
 import kotlinx.uuid.generateUUID
 
 class GcpChat(
+  private val provider: GCP, // TODO: use context receiver
   override val modelType: ModelType,
-  private val client: GcpClient,
 ) : Chat {
 
+  private val client = provider.defaultClient
+
+  override fun copy(modelType: ModelType) = GcpChat(provider, modelType)
+
   override suspend fun createChatCompletion(
     request: ChatCompletionRequest
   ): ChatCompletionResponse {

integrations/gcp/src/commonMain/kotlin/com/xebia/functional/xef/gcp/models/GcpCompletion.kt

@@ -1,7 +1,7 @@
 package com.xebia.functional.xef.gcp.models
 
 import com.xebia.functional.tokenizer.ModelType
-import com.xebia.functional.xef.gcp.GcpClient
+import com.xebia.functional.xef.gcp.GCP
 import com.xebia.functional.xef.llm.Completion
 import com.xebia.functional.xef.llm.models.text.CompletionChoice
 import com.xebia.functional.xef.llm.models.text.CompletionRequest
@@ -12,10 +12,14 @@ import kotlinx.uuid.UUID
 import kotlinx.uuid.generateUUID
 
 class GcpCompletion(
+  private val provider: GCP, // TODO: use context receiver
   override val modelType: ModelType,
-  private val client: GcpClient,
 ) : Completion {
 
+  private val client = provider.defaultClient
+
+  override fun copy(modelType: ModelType) = GcpCompletion(provider, modelType)
+
   override suspend fun createCompletion(request: CompletionRequest): CompletionResult {
     val response: String =
       client.promptMessage(

integrations/gcp/src/commonMain/kotlin/com/xebia/functional/xef/gcp/models/GcpEmbeddings.kt

@@ -1,6 +1,7 @@
 package com.xebia.functional.xef.gcp.models
 
 import com.xebia.functional.tokenizer.ModelType
+import com.xebia.functional.xef.gcp.GCP
 import com.xebia.functional.xef.gcp.GcpClient
 import com.xebia.functional.xef.llm.Embeddings
 import com.xebia.functional.xef.llm.models.embeddings.Embedding
@@ -9,10 +10,14 @@ import com.xebia.functional.xef.llm.models.embeddings.EmbeddingResult
 import com.xebia.functional.xef.llm.models.usage.Usage
 
 class GcpEmbeddings(
+  private val provider: GCP, // TODO: use context receiver
   override val modelType: ModelType,
-  private val client: GcpClient,
 ) : Embeddings {
 
+  private val client = provider.defaultClient
+
+  override fun copy(modelType: ModelType) = GcpEmbeddings(provider, modelType)
+
   override suspend fun createEmbeddings(request: EmbeddingRequest): EmbeddingResult {
     fun requestToEmbedding(it: GcpClient.EmbeddingPredictions): Embedding =
       Embedding(it.embeddings.values.map(Double::toFloat))

integrations/postgresql/src/test/kotlin/xef/PGVectorStoreSpec.kt

@@ -110,6 +110,9 @@ class PGVectorStoreSpec :
   })
 
 class TestLLM(override val modelType: ModelType = ModelType.ADA) : Chat, AutoCloseable {
+  override fun copy(modelType: ModelType) =
+    TestLLM(modelType)
+
   override fun tokensFromMessages(messages: List<Message>): Int = messages.map { calculateTokens(it) }.sum()
 
   private fun calculateTokens(message: Message): Int = message.content.split(" ").size + 2 // 2 is the role and name
@@ -145,6 +148,9 @@ private fun Embeddings.Companion.mock(
   }
 ): Embeddings =
   object : Embeddings {
+    override fun copy(modelType: ModelType): LLM {
+      throw NotImplementedError()
+    }
     override suspend fun embedDocuments(
       texts: List<String>,
       requestConfig: RequestConfig,

openai/src/commonMain/kotlin/com/xebia/functional/xef/conversation/llm/openai/OpenAI.kt

@@ -1,7 +1,9 @@
 package com.xebia.functional.xef.conversation.llm.openai
 
 import arrow.core.nonEmptyListOf
+import com.aallam.openai.api.exception.InvalidRequestException
 import com.aallam.openai.api.logging.LogLevel
+import com.aallam.openai.api.model.ModelId
 import com.aallam.openai.client.LoggingConfig
 import com.aallam.openai.client.OpenAI as OpenAIClient
 import com.aallam.openai.client.OpenAIHost
@@ -57,7 +59,7 @@ class OpenAI(internal var token: String? = null, internal var host: String? = nu
     }
   }
 
-  val defaultClient =
+  internal val defaultClient =
     OpenAIClient(
         host = getHost()?.let { OpenAIHost(it) } ?: OpenAIHost.OpenAI,
         token = getToken(),
@@ -66,51 +68,43 @@ class OpenAI(internal var token: String? = null, internal var host: String? = nu
       )
       .let { autoClose(it) }
 
-  val GPT_4 by lazy { autoClose(OpenAIChat(ModelType.GPT_4, defaultClient)) }
+  val GPT_4 by lazy { autoClose(OpenAIChat(this, ModelType.GPT_4)) }
 
   val GPT_4_0314 by lazy {
-    autoClose(OpenAIFunChat(ModelType.GPT_4_0314, defaultClient)) // legacy
+    autoClose(OpenAIFunChat(this, ModelType.GPT_4_0314)) // legacy
   }
 
-  val GPT_4_32K by lazy { autoClose(OpenAIChat(ModelType.GPT_4_32K, defaultClient)) }
+  val GPT_4_32K by lazy { autoClose(OpenAIChat(this, ModelType.GPT_4_32K)) }
 
-  val GPT_3_5_TURBO by lazy { autoClose(OpenAIChat(ModelType.GPT_3_5_TURBO, defaultClient)) }
+  val GPT_3_5_TURBO by lazy { autoClose(OpenAIChat(this, ModelType.GPT_3_5_TURBO)) }
 
-  val GPT_3_5_TURBO_16K by lazy {
-    autoClose(OpenAIChat(ModelType.GPT_3_5_TURBO_16_K, defaultClient))
-  }
+  val GPT_3_5_TURBO_16K by lazy { autoClose(OpenAIChat(this, ModelType.GPT_3_5_TURBO_16_K)) }
 
   val GPT_3_5_TURBO_FUNCTIONS by lazy {
-    autoClose(OpenAIFunChat(ModelType.GPT_3_5_TURBO_FUNCTIONS, defaultClient))
+    autoClose(OpenAIFunChat(this, ModelType.GPT_3_5_TURBO_FUNCTIONS))
   }
 
   val GPT_3_5_TURBO_0301 by lazy {
-    autoClose(OpenAIChat(ModelType.GPT_3_5_TURBO, defaultClient)) // legacy
+    autoClose(OpenAIChat(this, ModelType.GPT_3_5_TURBO)) // legacy
   }
 
-  val TEXT_DAVINCI_003 by lazy {
-    autoClose(OpenAICompletion(ModelType.TEXT_DAVINCI_003, defaultClient))
-  }
+  val TEXT_DAVINCI_003 by lazy { autoClose(OpenAICompletion(this, ModelType.TEXT_DAVINCI_003)) }
 
-  val TEXT_DAVINCI_002 by lazy {
-    autoClose(OpenAICompletion(ModelType.TEXT_DAVINCI_002, defaultClient))
-  }
+  val TEXT_DAVINCI_002 by lazy { autoClose(OpenAICompletion(this, ModelType.TEXT_DAVINCI_002)) }
 
   val TEXT_CURIE_001 by lazy {
-    autoClose(OpenAICompletion(ModelType.TEXT_SIMILARITY_CURIE_001, defaultClient))
+    autoClose(OpenAICompletion(this, ModelType.TEXT_SIMILARITY_CURIE_001))
   }
 
-  val TEXT_BABBAGE_001 by lazy {
-    autoClose(OpenAICompletion(ModelType.TEXT_BABBAGE_001, defaultClient))
-  }
+  val TEXT_BABBAGE_001 by lazy { autoClose(OpenAICompletion(this, ModelType.TEXT_BABBAGE_001)) }
 
-  val TEXT_ADA_001 by lazy { autoClose(OpenAICompletion(ModelType.TEXT_ADA_001, defaultClient)) }
+  val TEXT_ADA_001 by lazy { autoClose(OpenAICompletion(this, ModelType.TEXT_ADA_001)) }
 
   val TEXT_EMBEDDING_ADA_002 by lazy {
-    autoClose(OpenAIEmbeddings(ModelType.TEXT_EMBEDDING_ADA_002, defaultClient))
+    autoClose(OpenAIEmbeddings(this, ModelType.TEXT_EMBEDDING_ADA_002))
   }
 
-  val DALLE_2 by lazy { autoClose(OpenAIImages(ModelType.GPT_3_5_TURBO, defaultClient)) }
+  val DALLE_2 by lazy { autoClose(OpenAIImages(this, ModelType.GPT_3_5_TURBO)) }
 
   @JvmField val DEFAULT_CHAT = GPT_3_5_TURBO_16K
 
@@ -120,8 +114,8 @@ class OpenAI(internal var token: String? = null, internal var host: String? = nu
 
   @JvmField val DEFAULT_IMAGES = DALLE_2
 
-  fun supportedModels(): List<LLM> =
-    listOf(
+  fun supportedModels(): List<LLM> = // TODO: impl of abstract provider function
+  listOf(
       GPT_4,
       GPT_4_0314,
       GPT_4_32K,
@@ -138,6 +132,28 @@ class OpenAI(internal var token: String? = null, internal var host: String? = nu
       DALLE_2,
     )
 
+  suspend fun findModel(modelId: String): Any? { // TODO: impl of abstract provider function
+    val model =
+      try {
+        defaultClient.model(ModelId(modelId))
+      } catch (e: InvalidRequestException) {
+        when (e.error.detail?.code) {
+          "model_not_found" -> return null
+          else -> throw e
+        }
+      }
+    return ModelType.TODO(model.id.id)
+  }
+
+  suspend fun <T : LLM> spawnModel(
+    modelId: String,
+    baseModel: T
+  ): T { // TODO: impl of abstract provider function
+    if (findModel(modelId) == null) error("model not found")
+    return baseModel.copy(ModelType.FineTunedModel(modelId, baseModel = baseModel.modelType)) as? T
+      ?: error("${baseModel::class} does not follow contract to return the most specific type")
+  }
+
   companion object {
 
     @JvmField val FromEnvironment: OpenAI = OpenAI()

openai/src/commonMain/kotlin/com/xebia/functional/xef/conversation/llm/openai/models/OpenAIChat.kt

@@ -4,8 +4,8 @@ import com.aallam.openai.api.chat.ChatChoice
 import com.aallam.openai.api.chat.ChatMessage
 import com.aallam.openai.api.chat.chatCompletionRequest
 import com.aallam.openai.api.model.ModelId
-import com.aallam.openai.client.OpenAI
 import com.xebia.functional.tokenizer.ModelType
+import com.xebia.functional.xef.conversation.llm.openai.OpenAI
 import com.xebia.functional.xef.conversation.llm.openai.toInternal
 import com.xebia.functional.xef.conversation.llm.openai.toOpenAI
 import com.xebia.functional.xef.llm.Chat
@@ -14,10 +14,14 @@ import kotlinx.coroutines.flow.Flow
 import kotlinx.coroutines.flow.map
 
 class OpenAIChat(
+  private val provider: OpenAI, // TODO: use context receiver
   override val modelType: ModelType,
-  private val client: OpenAI,
 ) : Chat {
 
+  private val client = provider.defaultClient
+
+  override fun copy(modelType: ModelType) = OpenAIChat(provider, modelType)
+
   override suspend fun createChatCompletion(
     request: ChatCompletionRequest
   ): ChatCompletionResponse {

openai/src/commonMain/kotlin/com/xebia/functional/xef/conversation/llm/openai/models/OpenAICompletion.kt

@@ -4,19 +4,23 @@ import com.aallam.openai.api.LegacyOpenAI
 import com.aallam.openai.api.completion.Choice
 import com.aallam.openai.api.completion.completionRequest
 import com.aallam.openai.api.model.ModelId
-import com.aallam.openai.client.OpenAI
 import com.xebia.functional.tokenizer.ModelType
+import com.xebia.functional.xef.conversation.llm.openai.OpenAI
 import com.xebia.functional.xef.conversation.llm.openai.toInternal
 import com.xebia.functional.xef.llm.Completion
 import com.xebia.functional.xef.llm.models.text.CompletionChoice
 import com.xebia.functional.xef.llm.models.text.CompletionRequest
 import com.xebia.functional.xef.llm.models.text.CompletionResult
 
 class OpenAICompletion(
+  private val provider: OpenAI, // TODO: use context receiver
   override val modelType: ModelType,
-  private val client: OpenAI,
 ) : Completion {
 
+  private val client = provider.defaultClient
+
+  override fun copy(modelType: ModelType) = OpenAICompletion(provider, modelType)
+
   @OptIn(LegacyOpenAI::class)
   override suspend fun createCompletion(request: CompletionRequest): CompletionResult {
     fun toInternal(it: Choice): CompletionChoice =