jasonhedman · Chel-Tbl · May 21, 2026
diff --git a/docker-compose.yml b/docker-compose.yml
@@ -3,7 +3,7 @@ version: "3.8"
 services:
   # PostgreSQL Database
   postgres:
-    image: postgres:15
+    image: pgvector/pgvector:pg15
     container_name: toolkit-postgres
     environment:
       POSTGRES_USER: postgres
@@ -48,4 +48,4 @@ services:
 volumes:
   postgres_data:
   redis_data:
-  blob_data:
+  blob_data:
diff --git a/prisma/migrations/20260521000100_file_embeddings/migration.sql b/prisma/migrations/20260521000100_file_embeddings/migration.sql
@@ -0,0 +1,27 @@
+CREATE EXTENSION IF NOT EXISTS vector;
+
+CREATE TYPE "FileEmbeddingStatus" AS ENUM ('skipped', 'indexing', 'indexed', 'failed');
+
+ALTER TABLE "File"
+ADD COLUMN "embeddingStatus" "FileEmbeddingStatus" NOT NULL DEFAULT 'skipped';
+
+CREATE TABLE "FileEmbeddingChunk" (
+    "id" TEXT NOT NULL,
+    "fileId" TEXT NOT NULL,
+    "userId" TEXT NOT NULL,
+    "chunkIndex" INTEGER NOT NULL,
+    "content" TEXT NOT NULL,
+    "embedding" vector(384) NOT NULL,
+    "createdAt" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
+
+    CONSTRAINT "FileEmbeddingChunk_pkey" PRIMARY KEY ("id")
+);
+
+CREATE UNIQUE INDEX "FileEmbeddingChunk_fileId_chunkIndex_key" ON "FileEmbeddingChunk"("fileId", "chunkIndex");
+CREATE INDEX "FileEmbeddingChunk_fileId_idx" ON "FileEmbeddingChunk"("fileId");
+CREATE INDEX "FileEmbeddingChunk_userId_idx" ON "FileEmbeddingChunk"("userId");
+CREATE INDEX "FileEmbeddingChunk_embedding_idx" ON "FileEmbeddingChunk" USING ivfflat ("embedding" vector_cosine_ops) WITH (lists = 100);
+
+ALTER TABLE "FileEmbeddingChunk"
+ADD CONSTRAINT "FileEmbeddingChunk_fileId_fkey"
+FOREIGN KEY ("fileId") REFERENCES "File"("id") ON DELETE CASCADE ON UPDATE CASCADE;
diff --git a/prisma/schema.prisma b/prisma/schema.prisma
@@ -115,13 +115,37 @@ model Video {
 }
 
 model File {
-    id          String   @id @default(uuid())
-    userId      String
-    name        String
-    contentType String
-    url         String
-    createdAt   DateTime @default(now())
-    user        User     @relation(fields: [userId], references: [id], onDelete: Cascade)
+    id              String               @id @default(uuid())
+    userId          String
+    name            String
+    contentType     String
+    url             String
+    embeddingStatus FileEmbeddingStatus  @default(skipped)
+    createdAt       DateTime             @default(now())
+    user            User                 @relation(fields: [userId], references: [id], onDelete: Cascade)
+    embeddingChunks FileEmbeddingChunk[]
+}
+
+enum FileEmbeddingStatus {
+    skipped
+    indexing
+    indexed
+    failed
+}
+
+model FileEmbeddingChunk {
+    id         String                     @id @default(uuid())
+    fileId     String
+    userId     String
+    chunkIndex Int
+    content    String                     @db.Text
+    embedding  Unsupported("vector(384)")
+    createdAt  DateTime                   @default(now())
+    file       File                       @relation(fields: [fileId], references: [id], onDelete: Cascade)
+
+    @@unique([fileId, chunkIndex])
+    @@index([fileId])
+    @@index([userId])
 }
 
 model Image {
@@ -190,4 +214,4 @@ model Tool {
     toolkit     Toolkit  @relation(fields: [toolkitId], references: [id], onDelete: Cascade)
 
     @@id([id, toolkitId])
-}
+}
diff --git a/src/app/(general)/_components/chat/input/index.tsx b/src/app/(general)/_components/chat/input/index.tsx
@@ -96,6 +96,9 @@ const PureMultimodalInput: React.FC<Props> = ({
     "input",
     "",
   );
+  const [fileIndexingEnabled, setFileIndexingEnabled] = useLocalStorage<
+    boolean | null
+  >("file-indexing-enabled", null);
 
   useEffect(() => {
     if (textareaRef.current) {
@@ -121,6 +124,18 @@ const PureMultimodalInput: React.FC<Props> = ({
   const fileInputRef = useRef<HTMLInputElement>(null);
   const [uploadQueue, setUploadQueue] = useState<Array<string>>([]);
 
+  const requestFileIndexingPreference = useCallback(() => {
+    if (fileIndexingEnabled !== null) {
+      return fileIndexingEnabled;
+    }
+
+    const shouldIndex = window.confirm(
+      "Index uploaded files for retrieval? This stores searchable chunks in your database and enables the File Search toolkit.",
+    );
+    setFileIndexingEnabled(shouldIndex);
+    return shouldIndex;
+  }, [fileIndexingEnabled, setFileIndexingEnabled]);
+
   const supportsImages = selectedChatModel?.capabilities?.includes(
     LanguageModelCapability.Vision,
   );
@@ -199,9 +214,10 @@ const PureMultimodalInput: React.FC<Props> = ({
   ]);
 
   const uploadFile = useCallback(
-    async (file: File): Promise<Attachment | undefined> => {
+    async (file: File, indexFile: boolean): Promise<Attachment | undefined> => {
       const formData = new FormData();
       formData.append("file", file);
+      formData.append("indexFile", String(indexFile));
 
       try {
         const response = await fetch("/api/files/upload", {
@@ -235,11 +251,15 @@ const PureMultimodalInput: React.FC<Props> = ({
   const handleFileChange = useCallback(
     async (event: React.ChangeEvent<HTMLInputElement>) => {
       const files = Array.from(event.target.files ?? []);
+      const indexFiles =
+        files.length > 0 ? requestFileIndexingPreference() : false;
 
       setUploadQueue(files.map((file) => file.name));
 
       try {
-        const uploadPromises = files.map((file) => uploadFile(file));
+        const uploadPromises = files.map((file) =>
+          uploadFile(file, indexFiles),
+        );
         const uploadedAttachments = await Promise.all(uploadPromises);
         const successfullyUploadedAttachments = uploadedAttachments.filter(
           (attachment): attachment is Attachment => attachment !== undefined,
@@ -255,7 +275,7 @@ const PureMultimodalInput: React.FC<Props> = ({
         setUploadQueue([]);
       }
     },
-    [setAttachments, uploadFile],
+    [requestFileIndexingPreference, setAttachments, uploadFile],
   );
 
   useEffect(() => {
@@ -302,10 +322,11 @@ const PureMultimodalInput: React.FC<Props> = ({
           return;
         }
 
+        const indexFiles = requestFileIndexingPreference();
         setUploadQueue(imageFiles.map((file) => file.name));
 
         // Handle async upload in non-blocking way
-        Promise.all(imageFiles.map((file) => uploadFile(file)))
+        Promise.all(imageFiles.map((file) => uploadFile(file, indexFiles)))
           .then((uploadedAttachments) => {
             const successfullyUploadedAttachments = uploadedAttachments.filter(
               (attachment): attachment is Attachment =>
@@ -332,7 +353,7 @@ const PureMultimodalInput: React.FC<Props> = ({
           });
       }
     },
-    [supportsImages, setAttachments, uploadFile],
+    [requestFileIndexingPreference, supportsImages, setAttachments, uploadFile],
   );
 
   useEffect(() => {

diff --git a/src/app/(general)/account/components/tabs/attachments/table.tsx b/src/app/(general)/account/components/tabs/attachments/table.tsx
@@ -36,6 +36,7 @@ export function DataTableDemo() {
   const [visibleColumns, setVisibleColumns] = React.useState({
     name: true,
     contentType: true,
+    embeddingStatus: true,
     actions: true,
   });
 
@@ -184,6 +185,9 @@ export function DataTableDemo() {
               {visibleColumns.contentType && (
                 <TableHead>Content Type</TableHead>
               )}
+              {visibleColumns.embeddingStatus && (
+                <TableHead>Indexing</TableHead>
+              )}
               {visibleColumns.actions && (
                 <TableHead className="w-12"></TableHead>
               )}
@@ -237,6 +241,13 @@ export function DataTableDemo() {
                       <div className="capitalize">{attachment.contentType}</div>
                     </TableCell>
                   )}
+                  {visibleColumns.embeddingStatus && (
+                    <TableCell>
+                      <div className="capitalize">
+                        {attachment.embeddingStatus}
+                      </div>
+                    </TableCell>
+                  )}
                   {visibleColumns.actions && (
                     <TableCell>
                       <DropdownMenu>

diff --git a/src/app/api/files/upload/route.ts b/src/app/api/files/upload/route.ts
@@ -7,6 +7,7 @@ import { put } from "@vercel/blob";
 import { auth } from "@/server/auth";
 import { api } from "@/trpc/server";
 import { FILE_MAX_SIZE, IS_DEVELOPMENT } from "@/lib/constants";
+import { indexUploadedFile } from "@/server/rag/files";
 
 // Use Blob instead of File since File is not available in Node.js environment
 const FileSchema = z.object({
@@ -20,7 +21,7 @@ const FileSchema = z.object({
       (file) =>
         ["image/jpeg", "image/png", "application/pdf"].includes(file.type),
       {
-        message: "File type should be JPEG or PNG",
+        message: "File type should be JPEG, PNG, or PDF",
       },
     ),
 });
@@ -56,6 +57,7 @@ export async function POST(request: Request) {
 
     // Get filename from formData since Blob doesn't have name property
     const filename = (formData.get("file") as File).name;
+    const shouldIndexFile = formData.get("indexFile") === "true";
     const fileBuffer = await file.arrayBuffer();
 
     try {
@@ -71,8 +73,19 @@ export async function POST(request: Request) {
         name: filename,
         url: data.url,
         contentType: validatedFile.data.file.type,
+        embeddingStatus: shouldIndexFile ? "indexing" : "skipped",
       });
 
+      if (shouldIndexFile) {
+        file.embeddingStatus = await indexUploadedFile({
+          fileId: file.id,
+          userId: session.user.id,
+          name: filename,
+          contentType: validatedFile.data.file.type,
+          arrayBuffer: fileBuffer,
+        });
+      }
+
       if (IS_DEVELOPMENT) {
         // In development, use base64 data URLs so openrouter can access the file
         // We can't use the local blob storage because it's not accessible to openrouter

diff --git a/src/lib/rag/embeddings.ts b/src/lib/rag/embeddings.ts
@@ -0,0 +1,66 @@
+export const FILE_EMBEDDING_DIMENSIONS = 384;
+
+const MAX_TOKENS = 2_000;
+const MAX_CHARACTER_FEATURES = 6_000;
+
+export function embedText(text: string) {
+  const vector = new Array<number>(FILE_EMBEDDING_DIMENSIONS).fill(0);
+  const normalized = text.normalize("NFKC").toLowerCase();
+  const tokens = normalized.match(/[\p{L}\p{N}]+/gu)?.slice(0, MAX_TOKENS);
+
+  if (tokens?.length) {
+    tokens.forEach((token) => addFeature(vector, `w:${token}`, 1));
+
+    for (let index = 0; index < tokens.length - 1; index++) {
+      addFeature(vector, `b:${tokens[index]!} ${tokens[index + 1]!}`, 0.7);
+    }
+  }
+
+  const compactText = normalized.replace(/\s+/g, " ");
+  const characterFeatureCount = Math.min(
+    compactText.length - 2,
+    MAX_CHARACTER_FEATURES,
+  );
+
+  for (let index = 0; index < characterFeatureCount; index++) {
+    addFeature(vector, `c:${compactText.slice(index, index + 3)}`, 0.25);
+  }
+
+  normalizeVector(vector);
+  return vector;
+}
+
+export function toVectorLiteral(vector: number[]) {
+  return `[${vector.map((value) => Number(value.toFixed(6))).join(",")}]`;
+}
+
+function addFeature(vector: number[], feature: string, weight: number) {
+  const hash = hashString(feature);
+  const index = hash % FILE_EMBEDDING_DIMENSIONS;
+  const sign = hash & 1 ? 1 : -1;
+  vector[index] = (vector[index] ?? 0) + sign * weight;
+}
+
+function normalizeVector(vector: number[]) {
+  const norm = Math.sqrt(vector.reduce((sum, value) => sum + value ** 2, 0));
+
+  if (norm === 0) {
+    vector[0] = 1;
+    return;
+  }
+
+  for (let index = 0; index < vector.length; index++) {
+    vector[index] = vector[index]! / norm;
+  }
+}
+
+function hashString(value: string) {
+  let hash = 2166136261;
+
+  for (let index = 0; index < value.length; index++) {
+    hash ^= value.charCodeAt(index);
+    hash = Math.imul(hash, 16777619);
+  }
+
+  return hash >>> 0;
+}