entirely local embeddings logic

electron-src/data/database.ts

@@ -114,8 +114,8 @@ export class SQLDatabase extends BaseDatabase<MesssagesDatabase> {
       indexMap.set(texts[i], i);
     }
     results.sort((a, b) => {
-      const aIndex = indexMap.get(a.text);
-      const bIndex = indexMap.get(b.text);
+      const aIndex = indexMap.get(a.text!);
+      const bIndex = indexMap.get(b.text!);
       if (aIndex === undefined || bIndex === undefined) {
         return 0;
       }

electron-src/data/embeddings-database.ts

@@ -2,6 +2,7 @@ import type { DB as EmbeddingsDb } from "../../_generated/embeddings-db";
 import logger from "../utils/logger";
 import { embeddingsDbPath } from "../utils/constants";
 import BaseDatabase from "./base-database";
+import cosineSimilarity from "../semantic-search/vector-comparison";
 
 export class EmbeddingsDatabase extends BaseDatabase<EmbeddingsDb> {
   embeddingsCache: { text: string; embedding: Float32Array }[] = [];
@@ -14,7 +15,16 @@ export class EmbeddingsDatabase extends BaseDatabase<EmbeddingsDb> {
     return result[0].count as number;
   };
 
-  getAllEmbeddings = async () => {
+  calculateSimilarity = async (embedding: Float32Array) => {
+    const allEmbeddings = await this.getAllEmbeddings();
+    const similarities = allEmbeddings.map((e) => {
+      const similarity = cosineSimilarity(embedding!, e.embedding);
+      return { similarity, text: e.text };
+    });
+    similarities.sort((a, b) => b.similarity - a.similarity);
+    return similarities.slice(0, 100).map((l) => l.text!);
+  };
+  loadVectorsIntoMemory = async () => {
     if (this.embeddingsCache.length) {
       return this.embeddingsCache;
     }
@@ -32,6 +42,9 @@ export class EmbeddingsDatabase extends BaseDatabase<EmbeddingsDb> {
         ),
       };
     });
+  };
+  getAllEmbeddings = async () => {
+    await this.loadVectorsIntoMemory();
     return this.embeddingsCache;
   };
   getEmbeddingByText = async (text: string) => {

electron-src/esbuild.main.config.ts

@@ -6,6 +6,7 @@ const config: BuildOptions = {
   entryPoints: [
     path.resolve("electron-src/index.ts"),
     path.resolve("electron-src/workers/worker.ts"),
+    path.resolve("electron-src/workers/embeddings-worker.ts"),
     path.resolve("electron-src/utils/preload.ts"),
   ],
   bundle: true,

electron-src/semantic-search/semantic-search.ts

@@ -5,8 +5,6 @@ import { handleIpc } from "../ipc/ipc";
 import logger from "../utils/logger";
 import { BatchOpenAi, OPENAI_EMBEDDING_MODEL } from "./batch-utils";
 import pMap from "p-map";
-import embeddingsDb from "../data/embeddings-database";
-import cosineSimilarity from "./vector-comparison";
 import { uniqBy } from "lodash-es";
 
 export interface SemanticSearchMetadata {
@@ -82,9 +80,9 @@ export const createEmbeddings = async ({ openAiKey }: { openAiKey: string }) =>
   });
   const openai = new OpenAIApi(configuration);
 
-  await embeddingsDb.initialize();
+  await dbWorker.embeddingsWorker.initialize();
 
-  const existingText = await embeddingsDb.getAllText();
+  const existingText = await dbWorker.embeddingsWorker.getAllText();
   const set = new Set(existingText);
   numCompleted = existingText.length;
   const notParsed = messages.filter((m) => m.text && !set.has(m.text));
@@ -105,7 +103,7 @@ export const createEmbeddings = async ({ openAiKey }: { openAiKey: string }) =>
         try {
           logger.info(`Inserting ${itemEmbeddings.length} vectors`);
           const embeddings = itemEmbeddings.map((l) => ({ embedding: l.values, text: l.metadata.text }));
-          await embeddingsDb.insertEmbeddings(embeddings);
+          await dbWorker.embeddingsWorker.insertEmbeddings(embeddings);
           logger.info(`Inserted ${itemEmbeddings.length} vectors`);
           numCompleted += itemEmbeddings.length;
         } catch (e) {
@@ -134,7 +132,7 @@ interface SemanticQueryOpts {
 }
 
 export async function semanticQuery({ queryText, openAiKey }: SemanticQueryOpts) {
-  const existingEmbedding = await embeddingsDb.getEmbeddingByText(queryText);
+  const existingEmbedding = await dbWorker.embeddingsWorker.getEmbeddingByText(queryText);
   let floatEmbedding = existingEmbedding?.embedding;
 
   if (!existingEmbedding) {
@@ -153,17 +151,11 @@ export async function semanticQuery({ queryText, openAiKey }: SemanticQueryOpts)
       return [];
     }
     // save embedding
-    await embeddingsDb.insertEmbeddings([{ embedding, text: queryText }]);
+    await dbWorker.embeddingsWorker.insertEmbeddings([{ embedding, text: queryText }]);
     floatEmbedding = new Float32Array(embedding);
   }
 
-  const allEmbeddings = await embeddingsDb.getAllEmbeddings();
-  const similarities = allEmbeddings.map((e) => {
-    const similarity = cosineSimilarity(floatEmbedding!, e.embedding);
-    return { similarity, text: e.text };
-  });
-  similarities.sort((a, b) => b.similarity - a.similarity);
-  return similarities.slice(0, 100).map((l) => l.text!);
+  return dbWorker.embeddingsWorker.calculateSimilarity(floatEmbedding!);
 }
 
 handleIpc("createEmbeddings", async ({ openAiKey: openAiKey }) => {
@@ -178,7 +170,7 @@ handleIpc("getEmbeddingsCompleted", async () => {
 
 handleIpc("calculateSemanticSearchStatsEnhanced", async () => {
   const stats = await dbWorker.worker.calculateSemanticSearchStats();
-  const localDb = embeddingsDb;
+  const localDb = dbWorker.embeddingsWorker;
   try {
     await localDb.initialize();
     const count = await localDb.countEmbeddings();

electron-src/semantic-search/vector-comparison.ts

@@ -35,11 +35,27 @@ function dot(x: ArrayLike, y: ArrayLike) {
   }
   return sum;
 }
-function cosineSimilarity(x: ArrayLike, y: ArrayLike): number {
-  const a = dot(x, y);
-  const b = l2norm(x);
-  const c = l2norm(y);
-  return a / (b * c);
-}
+// function cosineSimilarity(x: ArrayLike, y: ArrayLike): number {
+//   const a = dot(x, y);
+//   const b = l2norm(x);
+//   const c = l2norm(y);
+//   return a / (b * c);
+// }
+export default function cosineSimilarity(vectorA: Float32Array, vectorB: Float32Array) {
+  const dimensionality = Math.min(vectorA.length, vectorB.length);
+  let dotAB = 0;
+  let dotA = 0;
+  let dotB = 0;
+  let dimension = 0;
+  while (dimension < dimensionality) {
+    const componentA = vectorA[dimension];
+    const componentB = vectorB[dimension];
+    dotAB += componentA * componentB;
+    dotA += componentA * componentA;
+    dotB += componentB * componentB;
+    dimension += 1;
+  }
 
-export default cosineSimilarity;
+  const magnitude = Math.sqrt(dotA * dotB);
+  return magnitude === 0 ? 0 : dotAB / magnitude;
+}

electron-src/workers/database-worker.ts

@@ -6,15 +6,22 @@ import { copyLatestDb, localDbExists } from "../data/db-file-utils";
 import isDev from "electron-is-dev";
 import { join } from "path";
 import logger from "../utils/logger";
+import type { EmbeddingsDatabase } from "../data/embeddings-database";
+import embeddingsDb from "../data/embeddings-database";
+
 type WorkerType<T> = {
   [P in keyof T]: T[P] extends (...args: infer A) => infer R ? (...args: A) => Promise<R> : never;
 };
 
 class DbWorker {
   worker!: WorkerType<SQLDatabase> | SQLDatabase;
+  embeddingsWorker!: WorkerType<EmbeddingsDatabase> | EmbeddingsDatabase;
 
   startWorker = async () => {
     const path = isDev ? "workers/worker.js" : join("..", "..", "..", "app.asar.unpacked", "worker.js");
+    const embeddingWorkerPath = isDev
+      ? "workers/embeddings-worker.js"
+      : join("..", "..", "..", "app.asar.unpacked", "embeddings-worker.js");
 
     this.worker = isDev
       ? db
@@ -23,6 +30,13 @@ class DbWorker {
             resourceLimits: { maxOldGenerationSizeMb: 32678, maxYoungGenerationSizeMb: 32678 },
           }),
         );
+    this.embeddingsWorker = isDev
+      ? embeddingsDb
+      : await spawn<WorkerType<EmbeddingsDatabase>>(
+          new Worker(embeddingWorkerPath, {
+            resourceLimits: { maxOldGenerationSizeMb: 32678, maxYoungGenerationSizeMb: 32678 },
+          }),
+        );
   };
 
   setupHandlers() {
@@ -40,6 +54,8 @@ class DbWorker {
         handleIpc(prop, dbElement as any);
       }
     }
+
+    handleIpc("loadVectorsIntoMemory", this.embeddingsWorker.loadVectorsIntoMemory);
   }
   isCopying = false;
   doesLocalDbCopyExist = async () => {

electron-src/workers/embeddings-worker.ts

@@ -0,0 +1,14 @@
+import type { EmbeddingsDatabase } from "../data/embeddings-database";
+import { expose } from "threads/worker";
+import embeddingsDb from "../data/embeddings-database";
+
+const exposed: Partial<Record<Partial<keyof EmbeddingsDatabase>, any>> = {};
+for (const property in embeddingsDb) {
+  const prop = property as keyof EmbeddingsDatabase;
+  const dbElement = embeddingsDb[prop];
+  if (typeof dbElement === "function") {
+    exposed[prop] = dbElement;
+  }
+}
+
+expose(exposed);

package.json

@@ -3,7 +3,7 @@
   "author": "JonLuca DeCaro <mimessage@jonlu.ca>",
   "main": "build/electron-src/index.js",
   "name": "MiMessage",
-  "version": "1.0.10",
+  "version": "1.0.11",
   "productName": "Mimessage",
   "description": "Apple Messages UI alternative, with export, search, and more.",
   "scripts": {
@@ -155,11 +155,16 @@
       {
         "from": "build/electron-src/workers/worker.js",
         "to": "app.asar.unpacked/worker.js"
+      },
+      {
+        "from": "build/electron-src/workers/embeddings-worker.js",
+        "to": "app.asar.unpacked/embeddings-worker.js"
       }
     ],
     "mac": {
       "binaries": [
-        "build/electron-src/workers/worker.js"
+        "build/electron-src/workers/worker.js",
+        "build/electron-src/workers/embeddings-worker.js"
       ],
       "target": {
         "target": "default",

src/components/global-search/GlobalSearch.tsx

@@ -8,8 +8,9 @@ import {
   useGlobalSearch,
   useGroupChatList,
   useHandleMap,
+  useLoadSemanticResultsIntoMemory,
 } from "../../hooks/dataHooks";
-import { LinearProgress } from "@mui/material";
+import { CircularProgress, LinearProgress } from "@mui/material";
 import { Button, Checkbox, FormControlLabel, FormGroup } from "@mui/material";
 
 import { Virtuoso } from "react-virtuoso";
@@ -29,6 +30,7 @@ import { DayPicker } from "react-day-picker";
 import Popover from "@mui/material/Popover";
 import { shallow } from "zustand/shallow";
 import { SemanticSearchInfo } from "../chat/OpenAiKey";
+import Backdrop from "@mui/material/Backdrop";
 
 const GloablSearchInput = () => {
   const globalSearch = useMimessage((state) => state.globalSearch);
@@ -190,28 +192,69 @@ const GroupChatFilter = () => {
   );
 };
 
+const ToggleSemanticSearch = () => {
+  const { mutateAsync, isLoading } = useLoadSemanticResultsIntoMemory();
+  const { openAiKey, setUseSemanticSearch, useSemanticSearch } = useMimessage(
+    (state) => ({
+      useSemanticSearch: state.useSemanticSearch,
+      setUseSemanticSearch: state.setUseSemanticSearch,
+      openAiKey: state.openAiKey,
+    }),
+    shallow,
+  );
+  return (
+    <>
+      {isLoading && (
+        <Backdrop open>
+          <Box
+            onClick={(e) => e.stopPropagation()}
+            sx={{ background: "#2c2c2c", maxWidth: 600, p: 2, m: 2 }}
+            display={"flex"}
+            flexDirection={"column"}
+          >
+            <Typography variant="h1" sx={{ color: "white" }}>
+              Loading Vectors into Memory
+            </Typography>
+            <Typography variant="h6" sx={{ color: "white" }}>
+              This takes ~2s per 100k messages
+            </Typography>
+            <CircularProgress />
+          </Box>
+        </Backdrop>
+      )}
+      <FormGroup>
+        <FormControlLabel
+          control={
+            <Checkbox
+              style={{
+                color: "white",
+              }}
+              checked={useSemanticSearch}
+              onChange={() => {
+                setUseSemanticSearch(!useSemanticSearch);
+                mutateAsync();
+              }}
+              disabled={!openAiKey}
+              title={openAiKey ? "" : "OpenAI Key Required"}
+            />
+          }
+          label="Use Semantic Search"
+        />
+      </FormGroup>
+    </>
+  );
+};
+
 const GlobalSearchFilter = () => {
   const { data: results } = useGlobalSearch();
   const count = results?.length || 0;
-  const {
-    openAiKey,
-    setUseSemanticSearch,
-    useSemanticSearch,
-    startDate,
-    setStartDate,
-    setEndDate,
-    endDate,
-    globalSearch,
-  } = useMimessage(
+  const { startDate, setStartDate, setEndDate, endDate, globalSearch } = useMimessage(
     (state) => ({
       startDate: state.startDate,
       endDate: state.endDate,
       setStartDate: state.setStartDate,
       setEndDate: state.setEndDate,
       globalSearch: state.globalSearch,
-      useSemanticSearch: state.useSemanticSearch,
-      setUseSemanticSearch: state.setUseSemanticSearch,
-      openAiKey: state.openAiKey,
     }),
     shallow,
   );
@@ -237,22 +280,7 @@ const GlobalSearchFilter = () => {
       <GroupChatFilter />
       <DateFilter selection={startDate} setSelection={setStartDate} text={"Start Date"} />
       <DateFilter selection={endDate} setSelection={setEndDate} text={"End Date"} />
-      <FormGroup>
-        <FormControlLabel
-          control={
-            <Checkbox
-              style={{
-                color: "white",
-              }}
-              checked={useSemanticSearch}
-              onChange={() => setUseSemanticSearch(!useSemanticSearch)}
-              disabled={!openAiKey}
-              title={openAiKey ? "" : "OpenAI Key Required"}
-            />
-          }
-          label="Use Semantic Search"
-        />
-      </FormGroup>
+      <ToggleSemanticSearch />
     </Box>
   );
 };

src/hooks/dataHooks.ts

@@ -396,7 +396,6 @@ export const useGlobalSearch = () => {
     }),
     shallow,
   );
-
   const handleMap = useContactToHandleMap();
   return useQuery<GlobalSearchResponse>(
     [
@@ -612,6 +611,12 @@ export const useOpenFileAtLocation = () => {
   });
 };
 
+export const useLoadSemanticResultsIntoMemory = () => {
+  return useMutation(["loadVectorsIntoMemory"], async () => {
+    await ipcRenderer.invoke("loadVectorsIntoMemory");
+  });
+};
+
 export const useAccessibilityPermissionsCheck = () => {
   return useQuery<boolean | null>(
     ["accessibility-permissions"],