feat: summary before doc query for better context

src/lib/ai/openaiWrapper.ts

@@ -76,21 +76,26 @@ const completionConfig = {
 }
 
 export async function answerMeDirect(arg: {context: string, userPrompt: string, initPrompt?:string}): Promise<string> {
-  const defaultPrompt = "Answer the question as truthfully as possible using the provided text, and if the answer is not contained within the text below, say \"I don't know\"\n\n"
+  // const defaultPrompt = "Answer the question as truthfully as possible using the provided text, and if the answer is not contained within the text below, say \"I don't know\"\n\n"
+  const defaultPrompt = `Use the following pieces of context to answer the users question.
+If you don't know the answer, just say that you don't know, don't try to make up an answer.
+----------------
+`
   const initPrompt = arg.initPrompt ?? defaultPrompt
-  const {context, userPrompt} = arg
-  const prompt = initPrompt
-    + 'Context:\n' + context + '\n\n'
-    + 'Q: ' + userPrompt + '\nA: ';
 
-  logger.debug(prompt)
+  const {context, userPrompt} = arg
+  // const prompt = initPrompt
+  //   + 'Context:\n' + context + '\n\n'
+  //   + 'Q: ' + userPrompt + '\nA: ';
+  //
+  // logger.debug(prompt)
 
   const response = await callWithRetry(() => getOpenAIAPI(Config.chatModel).createChatCompletion({
     ...completionConfig,
     model: Config.chatModel,
     messages:[
-      {role: 'system', content: context},
-      {role: 'user', content: initPrompt},
+      {role: 'system', content: initPrompt + context},
+      // {role: 'user', content: context},
       {role: 'user', content: userPrompt}
     ],
     // user: TODO: for tracking purposes

src/pages/DocumentQuery.vue

@@ -47,7 +47,7 @@
 
 <script setup lang="ts">
 import {ref, nextTick, onMounted, Ref, computed} from 'vue'
-import {performQna2} from 'src/lib/ai/answer'
+import {performQna2, performQna3, performSummarisation} from 'src/lib/ai/answer'
 import {createVectorStoreFromLargeContent} from 'src/lib/ai/largeDocQna'
 import {exportFile, Notify} from 'quasar'
 import {matCloudUpload} from '@quasar/extras/material-icons'
@@ -85,10 +85,13 @@ async function doit() {
 
     const vectorStore = await createVectorStoreFromLargeContent(text.value, (p)=>embedProgress.value=p)
 
+    const summary = await performSummarisation(text.value)
+    console.log(`SUMMARY ${summary}`)
+
     let idx = 0
     for (const question of questionStore.questions) {
       console.log(`QUESTION ${idx}: ${question}`)
-      const response = await performQna2(question, vectorStore)
+      const response = await performQna3(question, summary, vectorStore)
       answers.value[idx] = response ?? 'cannot answer'
       answerLoading.value[idx] = false
       console.log(`ANSWER ${idx}: ${response}`)

src/pages/MultiFilePage.vue

@@ -10,7 +10,7 @@
 import MultiFileManager from 'components/MultiFileManager.vue'
 import QuestionInputs from 'components/QuestionInputs.vue'
 import {computed, ref} from 'vue'
-import {performQna2} from 'src/lib/ai/answer'
+import {performQna2, performQna3, performSummarisation} from 'src/lib/ai/answer'
 import {exportFile, Notify} from 'quasar'
 import {useQuestionStore} from 'stores/questionStore'
 import {useMultiFileStore} from 'stores/multiFileStore'
@@ -34,7 +34,7 @@ async function doit() {
       for (const [fileIdx, file] of multiFileStore.documentInfo.entries()) {
         console.log(`QUESTION ${idx}: ${question}`)
         const vectorStore = useMultiFileStore().vectorStore
-        const response = await performQna2(question, vectorStore, d=>d.metadata['name'] === file.file.name)
+        const response = await performQna3(question, file.summary, vectorStore, d=>d.metadata['name'] === file.file.name)
         answers[idx][fileIdx] = response ?? 'cannot answer'
         console.log(`ANSWER ${idx}: ${response}`)
         console.log()

src/stores/multiFileStore.ts

@@ -7,13 +7,15 @@ import {getLangchainConfig} from 'src/lib/ai/config';
 import {embedsCache} from 'src/lib/ai/openaiWrapper';
 import {anyBufferToText, fileToText} from 'src/lib/ai/unstructured';
 import {RecursiveCharacterTextSplitter} from 'langchain/text_splitter';
+import {performSummarisation} from "src/lib/ai/answer";
 
 export interface DocumentInfo {
   name: string
   file?: File
   buffer?: Buffer
-  status: 'pending' | 'parsing' | 'processing' | 'ready' | 'error'
+  status: 'pending' | 'parsing' | 'processing' | 'ready' | 'error' | string
   progress?: number,
+  summary?: string
   // vectors?: MemoryVectorStore
 }
 
@@ -58,6 +60,12 @@ export const useMultiFileStore = defineStore('multiFile', {
           const vectorStore = this.vectorStore
           await vectorStore.addDocuments(docs) // TODO: deduplicate based on metadata?
 
+          // We also want a summary
+          // TODO: This could be in parallel of above?
+          pendingDocument.status = 'summarising'
+          const summary = await performSummarisation(text)
+          pendingDocument.summary = summary
+
           // Important to markRaw to avoid proxying the insides
           // pendingDocument.vectors = markRaw(vectorStore)
           // Update the status to 'ready' on successful processing