feat: make spec workflow works

Author: phodal

cocoa-core/src/main/kotlin/cc/unitmesh/rag/splitter/EncodingTokenizer.kt cocoa-core/src/main/kotlin/cc/unitmesh/nlp/embedding/EncodingTokenizer.kt

@@ -1,4 +1,4 @@
-package cc.unitmesh.rag.splitter
+package cc.unitmesh.nlp.embedding
 
 interface EncodingTokenizer {
     fun encode(text: String): List<Int>

cocoa-core/src/main/kotlin/cc/unitmesh/rag/splitter/OpenAiEncoding.kt cocoa-core/src/main/kotlin/cc/unitmesh/nlp/embedding/OpenAiEncoding.kt

@@ -1,4 +1,4 @@
-package cc.unitmesh.rag.splitter
+package cc.unitmesh.nlp.embedding
 
 import com.knuddels.jtokkit.Encodings
 import com.knuddels.jtokkit.api.Encoding

cocoa-core/src/main/kotlin/cc/unitmesh/rag/splitter/TokenTextSplitter.kt

@@ -17,6 +17,8 @@
  */
 package cc.unitmesh.rag.splitter
 
+import cc.unitmesh.nlp.embedding.EncodingTokenizer
+import cc.unitmesh.nlp.embedding.OpenAiEncoding
 import kotlin.math.max
 import kotlin.math.min
 

src/main/kotlin/cc/unitmesh/cf/domains/spec/SpecRelevantSearch.kt

@@ -2,6 +2,7 @@ package cc.unitmesh.cf.domains.spec
 
 import cc.unitmesh.nlp.embedding.Embedding
 import cc.unitmesh.nlp.embedding.EmbeddingProvider
+import cc.unitmesh.nlp.embedding.EncodingTokenizer
 import cc.unitmesh.rag.document.Document
 import cc.unitmesh.rag.retriever.EmbeddingStoreRetriever
 import cc.unitmesh.rag.splitter.MarkdownHeaderTextSplitter
@@ -15,40 +16,53 @@ class SpecRelevantSearch(val embeddingProvider: EmbeddingProvider) {
     private lateinit var vectorStoreRetriever: EmbeddingStoreRetriever
 
     // cached for performance
-    private val searchCache: MutableMap<String, List<String>> = mutableMapOf()
+    private val searchCache: MutableMap<String, List<SearchResult>> = mutableMapOf()
 
     init {
         val text = javaClass.getResourceAsStream("/be/specification.md")!!.bufferedReader().readText()
         val headersToSplitOn: List<Pair<String, String>> = listOf(
-            Pair("#", "Header 1"),
-            Pair("##", "Header 2"),
-            Pair("###", "Header 3"),
+            Pair("#", "H1"),
+            Pair("##", "H2"),
         )
 
         val documents = MarkdownHeaderTextSplitter(headersToSplitOn)
             .splitText(text)
 
-        val documentList = TokenTextSplitter(chunkSize = 384).apply(documents)
+        val documentList = documents.map {
+            val header = "${it.metadata["H1"]} > ${it.metadata["H2"]}"
+            val withHeader = it.copy(text = "$header ${it.text}")
+            TokenTextSplitter(chunkSize = 384).apply(listOf(withHeader)).first()
+        }
 
         val vectorStore: EmbeddingStore<Document> = InMemoryEmbeddingStore()
         val embeddings: List<Embedding> = documentList.map {
             embeddingProvider.embed(it.text)
         }
         vectorStore.addAll(embeddings, documentList)
 
-        this.vectorStoreRetriever = EmbeddingStoreRetriever(vectorStore)
+        this.vectorStoreRetriever = EmbeddingStoreRetriever(vectorStore, 5, 0.6)
     }
 
     // TODO: change to search engine
-    fun search(query: String): List<String> {
+    fun search(query: String): List<SearchResult> {
         if (searchCache.containsKey(query)) {
             return searchCache[query]!!
         }
 
         val queryEmbedding = embeddingProvider.embed(query)
         val similarDocuments = vectorStoreRetriever.retrieve(queryEmbedding)
-        val results = similarDocuments.map { it.embedded.text }
+        val results = similarDocuments.map {
+                SearchResult(
+                    source = it.embedded.metadata.toString(),
+                    content = it.embedded.text
+                )
+        }
         searchCache[query] = results
         return results
     }
-}
+}
+
+data class SearchResult(
+    val source: String,
+    val content: String
+)

src/main/kotlin/cc/unitmesh/cf/domains/spec/SpecWorkflow.kt

@@ -26,14 +26,18 @@ class SpecWorkflow : Workflow() {
         )
 
     override fun execute(prompt: StageContext, chatWebContext: ChatWebContext): Flowable<WorkflowResult> {
-        // TODO clarify user question, 如系统包含了这些规范，你需要哪些规范？
-        val specs = relevantSearch.search(chatWebContext.messages.last {
+        val question = chatWebContext.messages.last {
             it.role == LlmMsg.ChatRole.User.value
-        }.content)
+        }.content
+
+        // TODO clarify user question, 如系统包含了这些规范，你需要哪些规范？
+        val specs = relevantSearch.search(question)
 
         val userMsg = EXECUTE.format()
-            .replace("${'$'}{specs}", specs.joinToString("\n"))
-            .replace("${'$'}{question}", chatWebContext.messages[0].content)
+            .replace("${'$'}{specs}", specs.map {
+                "source: ${it.source} content: ${it.content}"
+            }.joinToString("\n"))
+            .replace("${'$'}{question}", question)
 
         val flowable = llmProvider.streamCompletion(listOf(
             LlmMsg.ChatMessage(LlmMsg.ChatRole.System, userMsg),
@@ -67,17 +71,27 @@ class SpecWorkflow : Workflow() {
             |
             |- 如果规范缺少对应的信息，你不要回答。
             |- 你必须回答用户的问题。
+            |- 请根据客户的问题，返回对应的规范，并返回对应的 source 相关信息。
+            |
             |
             |已有规范信息：
             |
             |```design
             |${'$'}{specs}
             |```
-            |用户的问题：
-            |${'$'}{question}
             |
-            |现在请你根据规范信息，回答用户的问题。
+            |示例：
+            |用户的问题：哪些规范包含了架构设计？
+            |回答：
+            |###
+            |
+            |出处：后端代码规范的命名规范章节 // 这里根据规范的 source 项信息，写出对应的来源
+            |// 这里，你需要返回是规范中的详细信息，而不是规范的标题。
+            |###
             |
+            |现在请你根据规范信息，回答用户的问题。
+            |用户的问题：
+            |${'$'}{question}
             |""".trimMargin()
         )
     }

src/main/kotlin/cc/unitmesh/cf/infrastructure/llms/embedding/SentenceTransformersEmbedding.kt

@@ -4,7 +4,7 @@ import org.springframework.stereotype.Component
 import cc.unitmesh.cf.STSemantic
 import cc.unitmesh.nlp.embedding.Embedding
 import cc.unitmesh.nlp.embedding.EmbeddingProvider
-import cc.unitmesh.rag.splitter.EncodingTokenizer
+import cc.unitmesh.nlp.embedding.EncodingTokenizer
 
 @Component
 class SentenceTransformersEmbedding : EmbeddingProvider, EncodingTokenizer {
@@ -26,8 +26,7 @@ class SentenceTransformersEmbedding : EmbeddingProvider, EncodingTokenizer {
 
     override fun decode(tokens: List<Int>): String {
         val map = tokens.map { it.toLong() }.toLongArray()
-        val output = tokenizer.decode(map)
         // output will be "[CLS] blog [SEP]" for input "blog", so we need to remove the first and last token
-        return output
+        return tokenizer.decode(map)
     }
 }