Merge with LLM Evaluation

backend/src/app.js

@@ -4,7 +4,7 @@ const dotenv = require("dotenv");
 const cors = require("cors");
 const session = require("express-session");
 const { initOpenAi } = require("./openai");
-const { initQAModel } = require("./documents");
+const { initQAModel, initPromptEvaluationModel } = require("./langchain");
 
 dotenv.config();
 
@@ -24,12 +24,15 @@ app.use(
   })
 );
 
-// initialise openai
+// main model for chat 
 initOpenAi();
 
-// initialise question answering llm
+// model for question answering of documenents
 initQAModel();
 
+// model for LLM prompt evaluation
+initPromptEvaluationModel();
+
 app.use(
   cors({
     credentials: true,

backend/src/defence.js

@@ -127,6 +127,7 @@ function detectTriggeredDefences(message, activeDefences) {
   return { reply: null, defenceInfo: defenceInfo };
 }
 
+
 module.exports = {
   activateDefence,
   deactivateDefence,

backend/src/langchain.js

@@ -0,0 +1,161 @@
+const { TextLoader } = require("langchain/document_loaders/fs/text");
+const { PDFLoader } = require("langchain/document_loaders/fs/pdf");
+const { CSVLoader } = require("langchain/document_loaders/fs/csv");
+const { DirectoryLoader } = require("langchain/document_loaders/fs/directory");
+const { RecursiveCharacterTextSplitter } = require("langchain/text_splitter");
+const { OpenAIEmbeddings } = require("langchain/embeddings/openai");
+const { MemoryVectorStore } = require("langchain/vectorstores/memory");
+const { ChatOpenAI } = require("langchain/chat_models/openai");
+const { RetrievalQAChain, LLMChain, SequentialChain } = require("langchain/chains");
+const { PromptTemplate } = require("langchain/prompts");
+const { OpenAI } = require("langchain/llms/openai");
+const { retrievalQATemplate, promptInjectionEvalTemplate, maliciousPromptTemplate } = require("./promptTemplates");
+
+// chain we use in question/answer request
+let qaChain = null;
+
+// chain we use in prompt evaluation request
+let promptEvaluationChain = null;
+
+// load the documents from filesystem
+async function getDocuments() {
+  const filePath = "resources/documents/";
+  const loader = new DirectoryLoader(filePath, {
+    ".pdf": (path) => new PDFLoader(path),
+    ".txt": (path) => new TextLoader(path),
+    ".csv": (path) => new CSVLoader(path),
+  });
+  const docs = await loader.load();
+
+  // split the documents into chunks
+  const textSplitter = new RecursiveCharacterTextSplitter({
+    chunkSize: 1000,
+    chunkOverlap: 0,
+  });
+  const splitDocs = await textSplitter.splitDocuments(docs);
+
+  console.debug(
+    "Documents loaded and split. Number of chunks: " + splitDocs.length
+  );
+
+  return splitDocs;
+}
+
+// QA Chain - ask the chat model a question about the documents
+async function initQAModel() {
+  // get the documents
+  const docs = await getDocuments();
+
+  // embed and store the splits
+  const embeddings = new OpenAIEmbeddings();
+  const vectorStore = await MemoryVectorStore.fromDocuments(
+    docs,
+    embeddings,
+  );
+  // initialise model
+  const model = new ChatOpenAI({
+    modelName: "gpt-4",
+    stream: true,
+  });
+
+  // prompt template for question and answering
+  const qaPrompt = PromptTemplate.fromTemplate(retrievalQATemplate);
+
+  // set chain to retrieval QA chain
+  qaChain = RetrievalQAChain.fromLLM(model, vectorStore.asRetriever(), {
+    prompt: qaPrompt,
+  });
+  console.debug("QA Retrieval chain initialised.");
+}
+
+
+// initialise the prompt evaluation model
+async function initPromptEvaluationModel() {
+
+  // create chain to detect prompt injection
+  const promptInjectionPrompt = PromptTemplate.fromTemplate(promptInjectionEvalTemplate);
+
+  const promptInjectionChain = new LLMChain({ 
+    llm: new OpenAI({ model: "gpt-3.5-turbo", temperature: 0 }), 
+    prompt: promptInjectionPrompt,
+    outputKey: "promptInjectionEval"
+  });
+
+  // create chain to detect malicious prompts
+  const maliciousInputPrompt = PromptTemplate.fromTemplate(maliciousPromptTemplate);
+  const maliciousInputChain = new LLMChain({ 
+    llm: new OpenAI({ model: "gpt-3.5-turbo", temperature: 0 }), 
+    prompt: maliciousInputPrompt,
+    outputKey: "maliciousInputEval"
+  });
+
+  promptEvaluationChain = new SequentialChain({
+    chains: [promptInjectionChain, maliciousInputChain],
+    inputVariables: ['prompt'],
+    outputVariables: ["promptInjectionEval", "maliciousInputEval"],
+  });
+  console.debug("Prompt evaluation chain initialised.");
+}
+
+
+// ask the question and return models answer
+async function queryDocuments(question) {
+  const response = await qaChain.call({
+    query: question,
+  });
+  console.debug("QA model response: " + response.text);
+  const result = {
+    reply: response.text,
+    questionAnswered: true,
+  };
+  return result;
+}
+
+// ask LLM whether the prompt is malicious
+async function queryPromptEvaluationModel(input) {
+  const response = await promptEvaluationChain.call({ prompt: input });
+
+  const promptInjectionEval = formatEvaluationOutput(response.promptInjectionEval);
+  const maliciousInputEval = formatEvaluationOutput(response.maliciousInputEval);
+
+  console.debug("Prompt injection eval: " + JSON.stringify(promptInjectionEval));
+  console.debug("Malicious input eval: " + JSON.stringify(maliciousInputEval));
+
+  // if both are malicious, combine reason
+  if (promptInjectionEval.isMalicious && maliciousInputEval.isMalicious) {
+    return { isMalicious: true, reason: `${promptInjectionEval.reason} & ${maliciousInputEval.reason}` };
+  } else if (promptInjectionEval.isMalicious) {
+      return { isMalicious: true, reason: promptInjectionEval.reason };
+  } else if (maliciousInputEval.isMalicious) {
+      return { isMalicious: true, reason: maliciousInputEval.reason };
+}
+return { isMalicious: false, reason: "" };
+}
+
+// format the evaluation model output. text should be a Yes or No answer followed by a reason
+function formatEvaluationOutput(response){
+  try {
+    // split response on first full stop or comma
+    const splitResponse = response.split(/\.|,/);
+    const answer = splitResponse[0].replace(/\W/g, '').toLowerCase();
+    const reason = splitResponse[1];
+    return {
+      isMalicious: answer === "yes",
+      reason: reason,
+    }
+
+  } catch (error) {
+    // in case the model does not respond in the format we have asked
+    console.error(error);
+    console.debug("Did not get a valid response from the prompt evaluation model. Original response: " + response.text);
+    return { isMalicious: false, reason: "" };
+  }
+}
+
+module.exports = {
+  getDocuments,
+  initQAModel,
+  initPromptEvaluationModel,
+  queryDocuments,
+  queryPromptEvaluationModel
+};

backend/src/openai.js

@@ -1,7 +1,7 @@
 const { Configuration, OpenAIApi } = require("openai");
 const { isDefenceActive } = require("./defence");
 const { sendEmail, getEmailWhitelist, isEmailInWhitelist } = require("./email");
-const { queryDocuments } = require("./documents");
+const { queryDocuments, queryPromptEvaluationModel } = require("./langchain");
 
 // OpenAI configuration
 let configuration = null;
@@ -33,15 +33,16 @@ const chatGptFunctions = [
   {
     name: "getEmailWhitelist",
     description:
-      "user asks who is on the email whitelist and the system replies with the list of emails. if the email whitelist defence is not active then user should be able to email anyone. ",
+      "user asks who is on the email whitelist and the system replies with the list of emails.",
     parameters: {
       type: "object",
       properties: {},
     },
   },
   {
     name: "askQuestion",
-    description: "Ask a question about the documents with company infomration",
+    description:
+      "Ask a question about the documents containing company information ",
     parameters: {
       type: "object",
       properties: {
@@ -104,7 +105,7 @@ async function chatGptCallFunction(functionCall, defenceInfo, session) {
           session
         );
       }
-    } else if (functionName == "getEmailWhitelist") {
+    } else if (functionName === "getEmailWhitelist") {
       response = getEmailWhitelist(
         isDefenceActive("EMAIL_WHITELIST", session.activeDefences)
       );
@@ -113,7 +114,7 @@ async function chatGptCallFunction(functionCall, defenceInfo, session) {
     if (functionName === "askQuestion") {
       console.debug("Asking question: " + params.question);
       // if asking a question, call the queryDocuments
-      response = await queryDocuments(params.question);
+      response = (await queryDocuments(params.question)).reply;
     }
 
     // add function call to chat
@@ -163,6 +164,20 @@ async function chatGptChatCompletion(session) {
 async function chatGptSendMessage(message, session) {
   // init defence info
   let defenceInfo = { triggeredDefences: [], blocked: false };
+
+  // evaluate the message for prompt injection
+  const evalPrompt = await queryPromptEvaluationModel(message);
+  if (evalPrompt.isMalicious) {
+    defenceInfo.triggeredDefences.push("LLM_EVALUATION");
+    if (isDefenceActive("LLM_EVALUATION", session.activeDefences)) {
+      console.debug("LLM evalutation defence active.");
+      defenceInfo.blocked = true;
+      const evalResponse =
+        "Message blocked by the malicious prompt evaluator." +
+        evalPrompt.reason;
+      return { reply: evalResponse, defenceInfo: defenceInfo };
+    }
+  }
   // add user message to chat
   session.chatHistory.push({ role: "user", content: message });
 
@@ -181,6 +196,7 @@ async function chatGptSendMessage(message, session) {
     session.chatHistory.push(functionCallReply.reply);
     // update the defence info
     defenceInfo = functionCallReply.defenceInfo;
+
     // get a new reply from ChatGPT now that the function has been called
     reply = await chatGptChatCompletion(session);
   }