Webui chat with x added

src/chain/chat-with-x.ts

@@ -0,0 +1,154 @@
+import { BaseLanguageModel } from "@langchain/core/language_models/base"
+import { Document } from "@langchain/core/documents"
+import {
+  ChatPromptTemplate,
+  MessagesPlaceholder,
+  PromptTemplate
+} from "@langchain/core/prompts"
+import { AIMessage, BaseMessage, HumanMessage } from "@langchain/core/messages"
+import { StringOutputParser } from "@langchain/core/output_parsers"
+import {
+  Runnable,
+  RunnableBranch,
+  RunnableLambda,
+  RunnableMap,
+  RunnableSequence
+} from "@langchain/core/runnables"
+type RetrievalChainInput = {
+  chat_history: string
+  question: string
+}
+
+const formatChatHistoryAsString = (history: BaseMessage[]) => {
+  return history
+    .map((message) => `${message._getType()}: ${message.content}`)
+    .join("\n")
+}
+
+export const formatDocs = (docs: Document[]) => {
+  return docs
+    .map((doc, i) => `<doc id='${i}'>${doc.pageContent}</doc>`)
+    .join("\n")
+}
+
+const serializeHistory = (input: any) => {
+  const chatHistory = input.chat_history || []
+  const convertedChatHistory = []
+  for (const message of chatHistory) {
+    if (message.human !== undefined) {
+      convertedChatHistory.push(new HumanMessage({ content: message.human }))
+    }
+    if (message["ai"] !== undefined) {
+      convertedChatHistory.push(new AIMessage({ content: message.ai }))
+    }
+  }
+  return convertedChatHistory
+}
+
+const createRetrieverChain = (
+  llm: BaseLanguageModel,
+  retriever: Runnable,
+  question_template: string
+) => {
+  const CONDENSE_QUESTION_PROMPT =
+    PromptTemplate.fromTemplate(question_template)
+  const condenseQuestionChain = RunnableSequence.from([
+    CONDENSE_QUESTION_PROMPT,
+    llm,
+    new StringOutputParser()
+  ]).withConfig({
+    runName: "CondenseQuestion"
+  })
+  const hasHistoryCheckFn = RunnableLambda.from(
+    (input: RetrievalChainInput) => input.chat_history.length > 0
+  ).withConfig({ runName: "HasChatHistoryCheck" })
+  const conversationChain = condenseQuestionChain.pipe(retriever).withConfig({
+    runName: "RetrievalChainWithHistory"
+  })
+  const basicRetrievalChain = RunnableLambda.from(
+    (input: RetrievalChainInput) => input.question
+  )
+    .withConfig({
+      runName: "Itemgetter:question"
+    })
+    .pipe(retriever)
+    .withConfig({ runName: "RetrievalChainWithNoHistory" })
+
+  return RunnableBranch.from([
+    [hasHistoryCheckFn, conversationChain],
+    basicRetrievalChain
+  ]).withConfig({
+    runName: "FindDocs"
+  })
+}
+
+export const createChatWithXChain = ({
+  llm,
+  question_template,
+  question_llm,
+  retriever,
+  response_template
+}: {
+  llm: BaseLanguageModel
+  question_llm: BaseLanguageModel
+  retriever: Runnable
+  question_template: string
+  response_template: string
+}) => {
+  const retrieverChain = createRetrieverChain(
+    question_llm,
+    retriever,
+    question_template
+  )
+  const context = RunnableMap.from({
+    context: RunnableSequence.from([
+      ({ question, chat_history }) => {
+        return {
+          question: question,
+          chat_history: formatChatHistoryAsString(chat_history)
+        }
+      },
+      retrieverChain,
+      RunnableLambda.from(formatDocs).withConfig({
+        runName: "FormatDocumentChunks"
+      })
+    ]),
+    question: RunnableLambda.from(
+      (input: RetrievalChainInput) => input.question
+    ).withConfig({
+      runName: "Itemgetter:question"
+    }),
+    chat_history: RunnableLambda.from(
+      (input: RetrievalChainInput) => input.chat_history
+    ).withConfig({
+      runName: "Itemgetter:chat_history"
+    })
+  }).withConfig({ tags: ["RetrieveDocs"] })
+  const prompt = ChatPromptTemplate.fromMessages([
+    ["system", response_template],
+    new MessagesPlaceholder("chat_history"),
+    ["human", "{question}"]
+  ])
+
+  const responseSynthesizerChain = RunnableSequence.from([
+    prompt,
+    llm,
+    new StringOutputParser()
+  ]).withConfig({
+    tags: ["GenerateResponse"]
+  })
+  return RunnableSequence.from([
+    {
+      question: RunnableLambda.from(
+        (input: RetrievalChainInput) => input.question
+      ).withConfig({
+        runName: "Itemgetter:question"
+      }),
+      chat_history: RunnableLambda.from(serializeHistory).withConfig({
+        runName: "SerializeHistory",
+      })
+    },
+    context,
+    responseSynthesizerChain
+  ])
+}

src/components/Option/Knowledge/SelectedKnwledge.tsx

@@ -12,7 +12,7 @@ export const SelectedKnowledge = ({ knowledge, onClose }: Props) => {
     <div className="mb-3 border flex justify-between items-center rounded-md p-2 dark:border-gray-600">
       <div className="flex flex-col gap-2">
         <div>
-          <h3 className="text-lg font-semibold dark:text-gray-100">
+          <h3 className="text-sm font-semibold dark:text-gray-100">
             {knowledge.title}
           </h3>
         </div>
@@ -20,7 +20,7 @@ export const SelectedKnowledge = ({ knowledge, onClose }: Props) => {
           {knowledge.source.map((source, index) => (
             <div
               key={index}
-              className="inline-flex gap-2 border rounded-md p-1 dark:border-gray-600 dark:text-gray-100">
+              className="inline-flex gap-2 text-xs border rounded-md p-1 dark:border-gray-600 dark:text-gray-100">
               <KnowledgeIcon type={source.type} className="w-4 h-4" />
               {source.filename}
             </div>

src/components/Option/Playground/PlaygroundForm.tsx

@@ -158,14 +158,7 @@ export const PlaygroundForm = ({ dropedFile }: Props) => {
   }
   return (
     <div className="px-3 pt-3 md:px-6 md:pt-6 bg-gray-50 dark:bg-[#262626] border rounded-t-xl  dark:border-gray-600">
-      {selectedKnowledge && (
+    
-        <SelectedKnowledge
-          onClose={() => {
-            setSelectedKnowledge(null)
-          }}
-          knowledge={selectedKnowledge}
-        />
-      )}
       <div
         className={`h-full rounded-md shadow relative ${
           form.values.image.length === 0 ? "hidden" : "block"

src/db/vector.ts

@@ -28,8 +28,8 @@ export class PageAssistVectorDb {
         } else {
           const data = result[id] as VectorData
           if (!data) {
-            console.log("Creating new vector")
+            console.log("Creating new vector", vector)
-            this.db.set({ [id]: { id, vectors: [vector] } }, () => {
+            this.db.set({ [id]: { id, vectors: vector } }, () => {
               if (chrome.runtime.lastError) {
                 reject(chrome.runtime.lastError)
               } else {

src/hooks/useMessageOption.tsx

@@ -1,8 +1,10 @@
 import React from "react"
 import { cleanUrl } from "~/libs/clean-url"
 import {
+  defaultEmbeddingModelForRag,
   geWebSearchFollowUpPrompt,
   getOllamaURL,
+  promptForRag,
   systemPromptForNonRagOption
 } from "~/services/ollama"
 import { type ChatHistory, type Message } from "~/store/option"
@@ -23,13 +25,16 @@ import { generateHistory } from "@/utils/generate-history"
 import { useTranslation } from "react-i18next"
 import { saveMessageOnError, saveMessageOnSuccess } from "./chat-helper"
 import { usePageAssist } from "@/context"
+import { OllamaEmbeddings } from "@langchain/community/embeddings/ollama"
+import { PageAssistVectorStore } from "@/libs/PageAssistVectorStore"
+import { formatDocs } from "@/chain/chat-with-x"
 
 export const useMessageOption = () => {
   const {
     controller: abortController,
     setController: setAbortController,
     messages,
-    setMessages,
+    setMessages
   } = usePageAssist()
   const {
     history,
@@ -502,6 +507,213 @@ export const useMessageOption = () => {
     }
   }
 
+  const ragMode = async (
+    message: string,
+    image: string,
+    isRegenerate: boolean,
+    messages: Message[],
+    history: ChatHistory,
+    signal: AbortSignal
+  ) => {
+    const url = await getOllamaURL()
+
+    const ollama = new ChatOllama({
+      model: selectedModel!,
+      baseUrl: cleanUrl(url)
+    })
+
+    let newMessage: Message[] = []
+    let generateMessageId = generateID()
+
+    if (!isRegenerate) {
+      newMessage = [
+        ...messages,
+        {
+          isBot: false,
+          name: "You",
+          message,
+          sources: [],
+          images: []
+        },
+        {
+          isBot: true,
+          name: selectedModel,
+          message: "▋",
+          sources: [],
+          id: generateMessageId
+        }
+      ]
+    } else {
+      newMessage = [
+        ...messages,
+        {
+          isBot: true,
+          name: selectedModel,
+          message: "▋",
+          sources: [],
+          id: generateMessageId
+        }
+      ]
+    }
+    setMessages(newMessage)
+    let fullText = ""
+    let contentToSave = ""
+
+    const embeddingModle = await defaultEmbeddingModelForRag()
+    const ollamaUrl = await getOllamaURL()
+    const ollamaEmbedding = new OllamaEmbeddings({
+      model: embeddingModle || selectedModel,
+      baseUrl: cleanUrl(ollamaUrl)
+    })
+
+    let vectorstore = await PageAssistVectorStore.fromExistingIndex(
+      ollamaEmbedding,
+      {
+        file_id: null,
+        knownledge_id: selectedKnowledge.id
+      }
+    )
+
+    try {
+      let query = message
+      const { ragPrompt: systemPrompt, ragQuestionPrompt: questionPrompt } =
+        await promptForRag()
+      if (newMessage.length > 2) {
+        const lastTenMessages = newMessage.slice(-10)
+        lastTenMessages.pop()
+        const chat_history = lastTenMessages
+          .map((message) => {
+            return `${message.isBot ? "Assistant: " : "Human: "}${message.message}`
+          })
+          .join("\n")
+        const promptForQuestion = questionPrompt
+          .replaceAll("{chat_history}", chat_history)
+          .replaceAll("{question}", message)
+        const questionOllama = new ChatOllama({
+          model: selectedModel!,
+          baseUrl: cleanUrl(url)
+        })
+        const response = await questionOllama.invoke(promptForQuestion)
+        query = response.content.toString()
+      }
+
+      const docs = await vectorstore.similaritySearch(query, 4)
+      const context = formatDocs(docs)
+      const source = docs.map((doc) => {
+        return {
+          name: doc?.metadata?.source || "untitled",
+          type: doc?.metadata?.type || "unknown",
+          url: ""
+        }
+      })
+      message = message.trim().replaceAll("\n", " ")
+
+      let humanMessage = new HumanMessage({
+        content: [
+          {
+            text: systemPrompt
+              .replace("{context}", context)
+              .replace("{question}", message),
+            type: "text"
+          }
+        ]
+      })
+
+      const applicationChatHistory = generateHistory(history)
+
+      const chunks = await ollama.stream(
+        [...applicationChatHistory, humanMessage],
+        {
+          signal: signal
+        }
+      )
+      let count = 0
+      for await (const chunk of chunks) {
+        contentToSave += chunk.content
+        fullText += chunk.content
+        if (count === 0) {
+          setIsProcessing(true)
+        }
+        setMessages((prev) => {
+          return prev.map((message) => {
+            if (message.id === generateMessageId) {
+              return {
+                ...message,
+                message: fullText.slice(0, -1) + "▋"
+              }
+            }
+            return message
+          })
+        })
+        count++
+      }
+      // update the message with the full text
+      setMessages((prev) => {
+        return prev.map((message) => {
+          if (message.id === generateMessageId) {
+            return {
+              ...message,
+              message: fullText,
+              sources: source
+            }
+          }
+          return message
+        })
+      })
+
+      setHistory([
+        ...history,
+        {
+          role: "user",
+          content: message,
+          image
+        },
+        {
+          role: "assistant",
+          content: fullText
+        }
+      ])
+
+      await saveMessageOnSuccess({
+        historyId,
+        setHistoryId,
+        isRegenerate,
+        selectedModel: selectedModel,
+        message,
+        image,
+        fullText,
+        source
+      })
+
+      setIsProcessing(false)
+      setStreaming(false)
+    } catch (e) {
+      const errorSave = await saveMessageOnError({
+        e,
+        botMessage: fullText,
+        history,
+        historyId,
+        image,
+        selectedModel,
+        setHistory,
+        setHistoryId,
+        userMessage: message,
+        isRegenerating: isRegenerate
+      })
+
+      if (!errorSave) {
+        notification.error({
+          message: t("error"),
+          description: e?.message || t("somethingWentWrong")
+        })
+      }
+      setIsProcessing(false)
+      setStreaming(false)
+    } finally {
+      setAbortController(null)
+    }
+  }
+
   const onSubmit = async ({
     message,
     image,
@@ -527,6 +739,16 @@ export const useMessageOption = () => {
       setAbortController(controller)
       signal = controller.signal
     }
+    if (selectedKnowledge) {
+      await ragMode(
+        message,
+        image,
+        isRegenerate,
+        chatHistory || messages,
+        memory || history,
+        signal
+      )
+    } else {
       if (webSearch) {
         await searchChatMode(
           message,
@@ -547,6 +769,7 @@ export const useMessageOption = () => {
         )
       }
     }
+  }
 
   const regenerateLastMessage = async () => {
     const isOk = validateBeforeSubmit()

src/libs/PageAssistVectorStore.ts

@@ -3,6 +3,7 @@ import { VectorStore } from "@langchain/core/vectorstores"
 import type { EmbeddingsInterface } from "@langchain/core/embeddings"
 import { Document } from "@langchain/core/documents"
 import { getVector, insertVector } from "@/db/vector"
+import { cp } from "fs"
 
 /**
  * Interface representing a vector in memory. It includes the content
@@ -116,8 +117,10 @@ export class PageAssistVectorStore extends VectorStore {
       })
       return filter(doc)
     }
-    const pgVector = await getVector(`vector:${this.knownledge_id}`)
+    const data = await getVector(`vector:${this.knownledge_id}`)
-    const filteredMemoryVectors = pgVector.vectors.filter(filterFunction)
+    const pgVector = [...data.vectors]
+    const filteredMemoryVectors = pgVector.filter(filterFunction)
+    console.log(filteredMemoryVectors)
     const searches = filteredMemoryVectors
       .map((vector, index) => ({
         similarity: this.similarity(query, vector.embedding),
@@ -125,7 +128,7 @@ export class PageAssistVectorStore extends VectorStore {
       }))
       .sort((a, b) => (a.similarity > b.similarity ? -1 : 0))
       .slice(0, k)
-
+    console.log(searches)
     const result: [Document, number][] = searches.map((search) => [
       new Document({
         metadata: filteredMemoryVectors[search.index].metadata,
@@ -133,7 +136,7 @@ export class PageAssistVectorStore extends VectorStore {
       }),
       search.similarity
     ])
-
+    console.log(result)
     return result
   }
 

src/loader/pdf-url.ts

@@ -40,7 +40,7 @@ export class PageAssistPDFUrlLoader
         .trim()
       documents.push({
         pageContent: text,
-        metadata: { source: this.name, page: i }
+        metadata: { source: this.name, page: i, type: "pdf" }
       })
     }