refactor: simplify chat history handling and update vector store inte…

…gration

src/app/api/ai/inner/route.ts

@@ -18,11 +18,7 @@ export async function POST(request: Request) {
 
   const reqJson = await request.json();
   const pastMessagesJsonString = reqJson.pastMessages;
-
-  const chatHistoryLines = parsePastMessagesToLines(
-    pastMessagesJsonString,
-    true
-  );
+  const chatHistoryLines = pastMessagesJsonString;
 
   console.log("");
   console.log("chatHistoryLines:");

src/app/api/ai/surface/route.ts

@@ -3,8 +3,8 @@ import { loadTridentSurfaceChain } from "@/utils/langchain/chains/surface";
 // using openai
 import { ChatOpenAI } from "@langchain/openai";
 import { OpenAIEmbeddings } from "@langchain/openai";
-import { parsePastMessagesToChatHistory } from "@/utils/trident/parsePastMessagesToChatHistory";
-import { BufferMemory } from "langchain/memory";
+import { parsePastMessagesToLines } from "@/utils/trident/parsePastMessagesToLines";
+import { MemoryVectorStore } from "langchain/vectorstores/memory";
 
 export async function POST(request: Request) {
   console.log("----- ----- -----");
@@ -13,14 +13,12 @@ export async function POST(request: Request) {
   const reqJson = await request.json();
   const query = reqJson.query;
   const pastMessagesJsonString = reqJson.pastMessages;
+  let chatHistoryLines = pastMessagesJsonString;
+  chatHistoryLines = chatHistoryLines + "\nHuman: " + query;
 
-  const chatHistory = parsePastMessagesToChatHistory(pastMessagesJsonString);
-
-  const memory = new BufferMemory({
-    returnMessages: true,
-    memoryKey: "history",
-    chatHistory,
-  });
+  console.log("");
+  console.log("chatHistoryLines:");
+  console.log(chatHistoryLines);
 
   let llm: ChatOpenAI;
   let embeddings: OpenAIEmbeddings;
@@ -41,27 +39,24 @@ export async function POST(request: Request) {
     embeddings = new OpenAIEmbeddings();
   }
 
+  const vectorStore = new MemoryVectorStore(embeddings);
+
   const surfaceChain = await loadTridentSurfaceChain({
     llm,
-    embeddings,
-    memory,
+    vectorStore,
   });
-  const surfaceResult = await surfaceChain.call({ input: query });
+  const surfaceResult = await surfaceChain.invoke({ input: chatHistoryLines });
 
   console.log("Human:", query);
-  console.log("AI:", surfaceResult.response);
+  console.log("AI:", surfaceResult.text);
   console.log("");
 
-  const history = await memory.chatHistory.getMessages();
-  // debug用
-  //console.log(history);
-
   console.log("----- end surface -----");
   console.log("----- ----- -----");
 
   return NextResponse.json({
     query: query,
-    surface: surfaceResult.response,
-    history: history,
+    surface: surfaceResult.text,
+    history: chatHistoryLines,
   });
 }

src/utils/langchain/chains/surface/index.ts

@@ -1,28 +1,16 @@
-import { Embeddings } from "@langchain/core/embeddings";
 import { loadTridentSurfacePrompt } from "./prompt";
 import { BaseLanguageModel } from "@langchain/core/language_models/base";
-import { BaseMemory } from "@langchain/core/memory";
 import { RunnableSequence } from "@langchain/core/runnables";
-import { BufferMemory } from "langchain/memory";
-import { ConversationChain } from "langchain/chains";
+import { VectorStore } from "@langchain/core/vectorstores";
 
 export const loadTridentSurfaceChain = async ({
-  embeddings,
   llm,
-  memory,
+  vectorStore,
 }: {
-  embeddings: Embeddings;
   llm: BaseLanguageModel;
-  memory?: BaseMemory;
-}): Promise<ConversationChain> => {
-  if (memory === undefined) {
-    memory = new BufferMemory();
-  }
-  const prompt = await loadTridentSurfacePrompt(embeddings);
-  const chain = new ConversationChain({
-    llm: llm,
-    prompt: prompt,
-    memory: memory,
-  });
+  vectorStore: VectorStore;
+}): Promise<RunnableSequence> => {
+  const prompt = await loadTridentSurfacePrompt(vectorStore);
+  const chain = RunnableSequence.from([prompt, llm]);
   return chain;
 };

src/utils/langchain/chains/surface/prompt.ts

@@ -1,7 +1,6 @@
-import { Embeddings } from "@langchain/core/embeddings";
 import { SemanticSimilarityExampleSelector } from "@langchain/core/example_selectors";
 import { FewShotPromptTemplate, PromptTemplate } from "@langchain/core/prompts";
-import { MemoryVectorStore } from "langchain/vectorstores/memory";
+import { VectorStore } from "@langchain/core/vectorstores";
 
 export const tridentSurfaceExampleList: Array<{
   input: string;
@@ -17,6 +16,11 @@ export const tridentSurfaceExampleList: Array<{
     output:
       "了解しました。OpenStreetMapのデータに基づいてニューヨーク市を表示する地図を作成しています。しばらくお待ちください……",
   },
+  {
+    input: "台東区を表示して",
+    output:
+      "了解しました。OpenStreetMapのデータに基づいて台東区を表示する地図を作成しています。しばらくお待ちください……",
+  },
   {
     input: "显示纽约地图",
     output: "知道了。我正在生成基于OpenStreetMap数据的纽约市地图。请稍等……",
@@ -42,10 +46,9 @@ You will always reply according to the following rules:
 
 ### Examples: ###`;
 
-export const loadTridentSurfacePrompt = async (embeddings: Embeddings) => {
-  const memoryVectorStore = new MemoryVectorStore(embeddings);
+export const loadTridentSurfacePrompt = async (vectorStore: VectorStore) => {
   const exampleSelector = new SemanticSimilarityExampleSelector({
-    vectorStore: memoryVectorStore,
+    vectorStore: vectorStore,
     k: 3,
     inputKeys: ["input"],
   });
@@ -69,10 +72,13 @@ AI:
     suffix: `
 ### Current conversation: ###
 
-{history}
-Human: {input}
-AI: `,
-    inputVariables: ["history", "input"],
+Human:
+{input}
+
+AI:
+`,
+    inputVariables: ["input"],
   });
+
   return dynamicPrompt;
 };