[DERCBOT-1609] Structure the LLM answer

Author: assouktim

bot/engine/src/main/kotlin/admin/bot/rag/BotRAGConfiguration.kt

@@ -36,6 +36,7 @@ data class BotRAGConfiguration(
     val llmSetting: LLMSetting? = null,
     val emSetting: EMSetting,
     val indexSessionId: String? = null,
+    @Deprecated("Replaced by LLM answer status")
     val noAnswerSentence: String,
     val noAnswerStoryId: String? = null,
     val documentsRequired: Boolean = true,

bot/engine/src/main/kotlin/engine/config/RAGAnswerHandler.kt

@@ -31,9 +31,9 @@ import ai.tock.bot.engine.action.SendSentenceWithFootnotes
 import ai.tock.bot.engine.dialog.Dialog
 import ai.tock.bot.engine.user.PlayerType
 import ai.tock.genai.orchestratorclient.requests.*
+import ai.tock.genai.orchestratorclient.responses.LLMAnswer
 import ai.tock.genai.orchestratorclient.responses.ObservabilityInfo
 import ai.tock.genai.orchestratorclient.responses.RAGResponse
-import ai.tock.genai.orchestratorclient.responses.TextWithFootnotes
 import ai.tock.genai.orchestratorclient.retrofit.GenAIOrchestratorBusinessError
 import ai.tock.genai.orchestratorclient.retrofit.GenAIOrchestratorValidationError
 import ai.tock.genai.orchestratorclient.services.RAGService
@@ -47,6 +47,8 @@ private val technicalErrorMessage = property(
     name = "tock_gen_ai_orchestrator_technical_error",
     defaultValue = "Technical error :( sorry!")
 
+// TODO MASS: Doit on enlever le champ NO_RAG_SENTENCE et la story associée de la config RAG (sudio Tock)?
+// TODO MASS: Il faudra changer tout les prompts
 
 object RAGAnswerHandler : AbstractProactiveAnswerHandler {
 
@@ -60,7 +62,7 @@ object RAGAnswerHandler : AbstractProactiveAnswerHandler {
             BotRepository.saveMetric(createMetric(MetricType.STORY_HANDLED))
 
             // Call RAG Api - Gen AI Orchestrator
-            val (answer, debug, noAnswerStory, observabilityInfo) = rag(this)
+            val (answer, footnotes, debug, noAnswerStory, observabilityInfo) = rag(this)
 
             // Add debug data if available and if debugging is enabled
             if (debug != null) {
@@ -75,14 +77,18 @@ object RAGAnswerHandler : AbstractProactiveAnswerHandler {
                 val modifiedObservabilityInfo = observabilityInfo?.let { updateObservabilityInfo(this, it) }
 
                 send(
-                    SendSentenceWithFootnotes(
-                        botId, connectorId, userId, text = answer.text, footnotes = answer.footnotes.map {
+                    action = SendSentenceWithFootnotes(
+                        playerId = botId,
+                        applicationId = connectorId,
+                        recipientId = userId,
+                        text = answer.answer,
+                        footnotes = footnotes?.map {
                             Footnote(
                                 it.identifier, it.title, it.url,
                                 if(action.metadata.sourceWithContent) it.content else null,
                                 it.score
                             )
-                        }.toMutableList(),
+                        }?.toMutableList() ?: mutableListOf<Footnote>(),
                         // modifiedObservabilityInfo includes the public langfuse URL if filled.
                         metadata = ActionMetadata(isGenAiRagAnswer = true, observabilityInfo = modifiedObservabilityInfo)
                     )
@@ -116,13 +122,13 @@ object RAGAnswerHandler : AbstractProactiveAnswerHandler {
     private fun ragStoryRedirection(botBus: BotBus, response: RAGResponse?): StoryDefinition? {
         return with(botBus) {
             botDefinition.ragConfiguration?.let { ragConfig ->
-                if (response?.answer?.text.equals(ragConfig.noAnswerSentence, ignoreCase = true)) {
+                if (response?.answer?.status.equals("not_found_in_context", ignoreCase = true)) {
                     // Save no answer metric
                     saveRagMetric(IndicatorValues.NO_ANSWER)
 
                     // Switch to no answer story if configured
                     if (!ragConfig.noAnswerStoryId.isNullOrBlank()) {
-                        logger.info { "The RAG response is equal to the configured no-answer sentence, so switch to the no-answer story." }
+                        logger.info { "Switch to the no-answer RAG story." }
                         getNoAnswerRAGStory(ragConfig)
                     } else null
                 } else {
@@ -221,7 +227,7 @@ object RAGAnswerHandler : AbstractProactiveAnswerHandler {
                 )
 
                 // Handle RAG response
-                return RAGResult(response?.answer, response?.debug, ragStoryRedirection(this, response), response?.observabilityInfo)
+                return RAGResult(response?.answer, response?.footnotes, response?.debug, ragStoryRedirection(this, response), response?.observabilityInfo)
             } catch (exc: Exception) {
                 logger.error { exc }
                 // Save failure metric
@@ -232,7 +238,7 @@ object RAGAnswerHandler : AbstractProactiveAnswerHandler {
                     RAGResult(noAnswerStory = getNoAnswerRAGStory(ragConfiguration))
                 }
                 else RAGResult(
-                    answer = TextWithFootnotes(text = technicalErrorMessage),
+                    answer = LLMAnswer(status="error", answer = technicalErrorMessage),
                     debug = when(exc) {
                         is GenAIOrchestratorBusinessError -> RAGError(exc.message, exc.error)
                         is GenAIOrchestratorValidationError -> RAGError(exc.message, exc.detail)
@@ -282,7 +288,8 @@ object RAGAnswerHandler : AbstractProactiveAnswerHandler {
  * Aggregation of RAG answer, debug and the no answer Story.
  */
 data class RAGResult(
-    val answer: TextWithFootnotes? = null,
+    val answer: LLMAnswer? = null,
+    val footnotes: List<ai.tock.genai.orchestratorclient.responses.Footnote>? = null,
     val debug: Any? = null,
     val noAnswerStory: StoryDefinition? = null,
     val observabilityInfo: ObservabilityInfo? = null,

gen-ai/orchestrator-client/src/main/kotlin/ai/tock/genai/orchestratorclient/responses/RAGResponse.kt

@@ -17,7 +17,8 @@
 package ai.tock.genai.orchestratorclient.responses
 
 data class RAGResponse(
-    val answer: TextWithFootnotes,
+    val answer: LLMAnswer,
+    val footnotes: List<Footnote> = emptyList(),
     val debug: Any? = null,
     val observabilityInfo: ObservabilityInfo? = null,
 )

gen-ai/orchestrator-client/src/main/kotlin/ai/tock/genai/orchestratorclient/responses/models.kt

@@ -17,9 +17,18 @@
 package ai.tock.genai.orchestratorclient.responses
 
 
-data class TextWithFootnotes(
-    val text: String,
-    val footnotes: List<Footnote> = emptyList(),
+data class ChunkSentences(
+    val chunk: Int,
+    val sentences: List<String> = emptyList(),
+    val reason: String? = null,
+)
+
+data class LLMAnswer(
+    val status: String,
+    val answer: String,
+    val topic: String? = null,
+    val suggestedTopics: List<String>? = null,
+    val context: List<ChunkSentences>? = null,
 )
 
 data class Footnote(

gen-ai/orchestrator-server/src/main/python/server/src/gen_ai_orchestrator/models/rag/rag_models.py

@@ -52,16 +52,17 @@ class Footnote(Source):
 
     identifier: str = Field(description='Footnote identifier', examples=['1'])
 
-
-class TextWithFootnotes(BaseModel):
-    """Text with its footnotes. Used for RAG response"""
-
-    text: str = Field(
-        description='Text with footnotes used to list outside sources',
-        examples=['This is page content [1], and this is more content [2]'],
-    )
-    footnotes: set[Footnote] = Field(description='Set of footnotes')
-
+class ChunkSentences(BaseModel):
+    chunk: int
+    sentences: List[str] = []
+    reason: Optional[str] = None
+
+class LLMAnswer(BaseModel):
+    status: Optional[str] = None
+    answer: Optional[str] = None
+    topic: Optional[str] = None
+    suggested_topics: Optional[List[str]] = None
+    context: Optional[List[ChunkSentences]] = None
 
 @unique
 class ChatMessageType(str, Enum):
@@ -154,4 +155,4 @@ class RAGDebugData(QADebugData):
             'Question: Hello, how to plan a trip to Morocco ?. Answer in French.'
         ],
     )
-    answer: str = Field(description='The RAG answer.')
+    answer: LLMAnswer = Field(description='The RAG answer.')

gen-ai/orchestrator-server/src/main/python/server/src/gen_ai_orchestrator/routers/responses/responses.py

@@ -25,9 +25,8 @@
     ErrorInfo,
 )
 from gen_ai_orchestrator.models.llm.llm_provider import LLMProvider
-from gen_ai_orchestrator.models.rag.rag_models import Source, TextWithFootnotes
+from gen_ai_orchestrator.models.rag.rag_models import Source, LLMAnswer, Footnote
 from gen_ai_orchestrator.models.observability.observability_provider import ObservabilityProvider
-from gen_ai_orchestrator.models.rag.rag_models import TextWithFootnotes
 from gen_ai_orchestrator.models.vector_stores.vectore_store_provider import VectorStoreProvider
 
 
@@ -122,9 +121,10 @@ class ObservabilityInfo(BaseModel):
 class RAGResponse(BaseModel):
     """The RAG response model"""
 
-    answer: TextWithFootnotes = Field(
-        description='The RAG answer, with outside sources.'
+    answer: Optional[LLMAnswer] = Field(
+        description='The RAG answer'
     )
+    footnotes: set[Footnote] = Field(description='Set of footnotes')
     debug: Optional[Any] = Field(
         description='Debug data',
         examples=[{'action': 'retrieve', 'result': 'OK', 'errors': []}],

gen-ai/orchestrator-server/src/main/python/server/src/gen_ai_orchestrator/services/langchain/factories/vector_stores/open_search_factory.py

@@ -61,7 +61,7 @@ def get_vector_store(self, async_mode: Optional[bool] = True) -> OpenSearchVecto
             # By default, is the same as host. If set to False, it will not verify hostname on certificate
             ssl_assert_hostname=self.setting.host if is_prod_environment else False,
             ssl_show_warn=is_prod_environment,
-            index_name=self.index_name,
+            index_name="ns-salledesmarchs-bot-sdm-session-f13384dd-084c-4414-bfbf-ad1f51e81490",
             embedding_function=self.embedding_function,
             timeout=application_settings.vector_store_timeout,
         )