feat: file attachment

core/src/browser/extension.ts

@@ -11,6 +11,8 @@ export enum ExtensionTypeEnum {
   HuggingFace = 'huggingFace',
   Engine = 'engine',
   Hardware = 'hardware',
+  RAG = 'rag',
+  VectorDB = 'vectorDB',
 }
 
 export interface ExtensionType {

core/src/browser/extensions/engines/AIEngine.ts

@@ -182,6 +182,7 @@ export interface SessionInfo {
   port: number // llama-server output port (corrected from portid)
   model_id: string //name of the model
   model_path: string // path of the loaded model
+  is_embedding: boolean
   api_key: string
   mmproj_path?: string
 }

core/src/browser/extensions/index.ts

@@ -23,3 +23,8 @@ export { MCPExtension } from './mcp'
  * Base AI Engines.
  */
 export * from './engines'
+
+export { RAGExtension, RAG_INTERNAL_SERVER } from './rag'
+export type { AttachmentInput, IngestAttachmentsResult } from './rag'
+export { VectorDBExtension } from './vector-db'
+export type { SearchMode, VectorDBStatus, VectorChunkInput, VectorSearchResult, AttachmentFileInfo, VectorDBFileInput, VectorDBIngestOptions } from './vector-db'

core/src/browser/extensions/rag.ts

@@ -0,0 +1,36 @@
+import { BaseExtension, ExtensionTypeEnum } from '../extension'
+import type { MCPTool, MCPToolCallResult } from '../../types'
+import type { AttachmentFileInfo } from './vector-db'
+
+export interface AttachmentInput {
+  path: string
+  name?: string
+  type?: string
+  size?: number
+}
+
+export interface IngestAttachmentsResult {
+  filesProcessed: number
+  chunksInserted: number
+  files: AttachmentFileInfo[]
+}
+
+export const RAG_INTERNAL_SERVER = 'rag-internal'
+
+/**
+ * RAG extension base: exposes RAG tools and orchestration API.
+ */
+export abstract class RAGExtension extends BaseExtension {
+  type(): ExtensionTypeEnum | undefined {
+    return ExtensionTypeEnum.RAG
+  }
+
+  abstract getTools(): Promise<MCPTool[]>
+  /**
+   * Lightweight list of tool names for quick routing/lookup.
+   */
+  abstract getToolNames(): Promise<string[]>
+  abstract callTool(toolName: string, args: Record<string, unknown>): Promise<MCPToolCallResult>
+
+  abstract ingestAttachments(threadId: string, files: AttachmentInput[]): Promise<IngestAttachmentsResult>
+}

core/src/browser/extensions/vector-db.ts

@@ -0,0 +1,82 @@
+import { BaseExtension, ExtensionTypeEnum } from '../extension'
+
+export type SearchMode = 'auto' | 'ann' | 'linear'
+
+export interface VectorDBStatus {
+  ann_available: boolean
+}
+
+export interface VectorChunkInput {
+  text: string
+  embedding: number[]
+}
+
+export interface VectorSearchResult {
+  id: string
+  text: string
+  score?: number
+  file_id: string
+  chunk_file_order: number
+}
+
+export interface AttachmentFileInfo {
+  id: string
+  name?: string
+  path?: string
+  type?: string
+  size?: number
+  chunk_count: number
+}
+
+// High-level input types for file ingestion
+export interface VectorDBFileInput {
+  path: string
+  name?: string
+  type?: string
+  size?: number
+}
+
+export interface VectorDBIngestOptions {
+  chunkSize: number
+  chunkOverlap: number
+}
+
+/**
+ * Vector DB extension base: abstraction over local vector storage and search.
+ */
+export abstract class VectorDBExtension extends BaseExtension {
+  type(): ExtensionTypeEnum | undefined {
+    return ExtensionTypeEnum.VectorDB
+  }
+
+  abstract getStatus(): Promise<VectorDBStatus>
+  abstract createCollection(threadId: string, dimension: number): Promise<void>
+  abstract insertChunks(
+    threadId: string,
+    fileId: string,
+    chunks: VectorChunkInput[]
+  ): Promise<void>
+  abstract ingestFile(
+    threadId: string,
+    file: VectorDBFileInput,
+    opts: VectorDBIngestOptions
+  ): Promise<AttachmentFileInfo>
+  abstract searchCollection(
+    threadId: string,
+    query_embedding: number[],
+    limit: number,
+    threshold: number,
+    mode?: SearchMode,
+    fileIds?: string[]
+  ): Promise<VectorSearchResult[]>
+  abstract deleteChunks(threadId: string, ids: string[]): Promise<void>
+  abstract deleteFile(threadId: string, fileId: string): Promise<void>
+  abstract deleteCollection(threadId: string): Promise<void>
+  abstract listAttachments(threadId: string, limit?: number): Promise<AttachmentFileInfo[]>
+  abstract getChunks(
+    threadId: string,
+    fileId: string,
+    startOrder: number,
+    endOrder: number
+  ): Promise<VectorSearchResult[]>
+}

core/src/types/setting/settingComponent.ts

@@ -12,6 +12,8 @@ export type SettingComponentProps = {
   extensionName?: string
   requireModelReload?: boolean
   configType?: ConfigType
+  titleKey?: string
+  descriptionKey?: string
 }
 
 export type ConfigType = 'runtime' | 'setting'

extensions-web/src/jan-provider-web/provider.ts

@@ -45,7 +45,7 @@ export default class JanProviderWeb extends AIEngine {
   // Verify Jan models capabilities in localStorage
   private validateJanModelsLocalStorage() {
     try {
-      console.log("Validating Jan models in localStorage...")
+      console.log('Validating Jan models in localStorage...')
       const storageKey = 'model-provider'
       const data = localStorage.getItem(storageKey)
       if (!data) return
@@ -60,9 +60,14 @@ export default class JanProviderWeb extends AIEngine {
         if (provider.provider === 'jan' && provider.models) {
           for (const model of provider.models) {
             console.log(`Checking Jan model: ${model.id}`, model.capabilities)
-            if (JSON.stringify(model.capabilities) !== JSON.stringify(JAN_MODEL_CAPABILITIES)) {
+            if (
+              JSON.stringify(model.capabilities) !==
+              JSON.stringify(JAN_MODEL_CAPABILITIES)
+            ) {
               hasInvalidModel = true
-              console.log(`Found invalid Jan model: ${model.id}, clearing localStorage`)
+              console.log(
+                `Found invalid Jan model: ${model.id}, clearing localStorage`
+              )
               break
             }
           }
@@ -79,9 +84,17 @@ export default class JanProviderWeb extends AIEngine {
         // If still present, try setting to empty state
         if (afterRemoval) {
           // Try alternative clearing method
-          localStorage.setItem(storageKey, JSON.stringify({ state: { providers: [] }, version: parsed.version || 3 }))
+          localStorage.setItem(
+            storageKey,
+            JSON.stringify({
+              state: { providers: [] },
+              version: parsed.version || 3,
+            })
+          )
         }
-        console.log('Cleared model-provider from localStorage due to invalid Jan capabilities')
+        console.log(
+          'Cleared model-provider from localStorage due to invalid Jan capabilities'
+        )
         // Force a page reload to ensure clean state
         window.location.reload()
       }
@@ -159,6 +172,7 @@ export default class JanProviderWeb extends AIEngine {
         port: 443, // HTTPS port
         model_id: modelId,
         model_path: `remote:${modelId}`, // Indicate this is a remote model
+        is_embedding: false, // assume false here, TODO: might need further implementation
         api_key: '', // API key handled by auth service
       }
 
@@ -193,8 +207,12 @@ export default class JanProviderWeb extends AIEngine {
       console.error(`Failed to unload Jan session ${sessionId}:`, error)
       return {
         success: false,
-        error: error instanceof ApiError ? error.message :
-               error instanceof Error ? error.message : 'Unknown error',
+        error:
+          error instanceof ApiError
+            ? error.message
+            : error instanceof Error
+              ? error.message
+              : 'Unknown error',
       }
     }
   }

extensions/llamacpp-extension/src/index.ts

@@ -333,14 +333,12 @@ export default class llamacpp_extension extends AIEngine {
           )
           // Clear the invalid stored preference
           this.clearStoredBackendType()
-          bestAvailableBackendString = await this.determineBestBackend(
-            version_backends
-          )
+          bestAvailableBackendString =
+            await this.determineBestBackend(version_backends)
         }
       } else {
-        bestAvailableBackendString = await this.determineBestBackend(
-          version_backends
-        )
+        bestAvailableBackendString =
+          await this.determineBestBackend(version_backends)
       }
 
       let settings = structuredClone(SETTINGS)
@@ -1530,17 +1528,41 @@ export default class llamacpp_extension extends AIEngine {
 
     if (
       this.autoUnload &&
+      !isEmbedding &&
       (loadedModels.length > 0 || otherLoadingPromises.length > 0)
     ) {
       // Wait for OTHER loading models to finish, then unload everything
       if (otherLoadingPromises.length > 0) {
         await Promise.all(otherLoadingPromises)
       }
 
-      // Now unload all loaded models
+      // Now unload all loaded Text models excluding embedding models
       const allLoadedModels = await this.getLoadedModels()
       if (allLoadedModels.length > 0) {
-        await Promise.all(allLoadedModels.map((model) => this.unload(model)))
+        const sessionInfos: (SessionInfo | null)[] = await Promise.all(
+          allLoadedModels.map(async (modelId) => {
+            try {
+              return await this.findSessionByModel(modelId)
+            } catch (e) {
+              logger.warn(`Unable to find session for model "${modelId}": ${e}`)
+              return null // treat as “not‑eligible for unload”
+            }
+          })
+        )
+
+        logger.info(JSON.stringify(sessionInfos))
+
+        const nonEmbeddingModels: string[] = sessionInfos
+          .filter(
+            (s): s is SessionInfo => s !== null && s.is_embedding === false
+          )
+          .map((s) => s.model_id)
+
+        if (nonEmbeddingModels.length > 0) {
+          await Promise.all(
+            nonEmbeddingModels.map((modelId) => this.unload(modelId))
+          )
+        }
       }
     }
     const args: string[] = []
@@ -1638,7 +1660,7 @@ export default class llamacpp_extension extends AIEngine {
     if (cfg.no_kv_offload) args.push('--no-kv-offload')
     if (isEmbedding) {
       args.push('--embedding')
-      args.push('--pooling mean')
+      args.push('--pooling', 'mean')
     } else {
       if (cfg.ctx_size > 0) args.push('--ctx-size', String(cfg.ctx_size))
       if (cfg.n_predict > 0) args.push('--n-predict', String(cfg.n_predict))
@@ -1677,6 +1699,7 @@ export default class llamacpp_extension extends AIEngine {
           libraryPath,
           args,
           envs,
+          isEmbedding,
         }
       )
       return sInfo
@@ -2083,6 +2106,7 @@ export default class llamacpp_extension extends AIEngine {
   }
 
   async embed(text: string[]): Promise<EmbeddingResponse> {
+    // Ensure the sentence-transformer model is present
     let sInfo = await this.findSessionByModel('sentence-transformer-mini')
     if (!sInfo) {
       const downloadedModelList = await this.list()
@@ -2096,30 +2120,45 @@ export default class llamacpp_extension extends AIEngine {
             'https://huggingface.co/second-state/All-MiniLM-L6-v2-Embedding-GGUF/resolve/main/all-MiniLM-L6-v2-ggml-model-f16.gguf?download=true',
         })
       }
-      sInfo = await this.load('sentence-transformer-mini')
+      // Load specifically in embedding mode
+      sInfo = await this.load('sentence-transformer-mini', undefined, true)
     }
-    const baseUrl = `http://localhost:${sInfo.port}/v1/embeddings`
-    const headers = {
-      'Content-Type': 'application/json',
-      'Authorization': `Bearer ${sInfo.api_key}`,
+
+    const attemptRequest = async (session: SessionInfo) => {
+      const baseUrl = `http://localhost:${session.port}/v1/embeddings`
+      const headers = {
+        'Content-Type': 'application/json',
+        'Authorization': `Bearer ${session.api_key}`,
+      }
+      const body = JSON.stringify({
+        input: text,
+        model: session.model_id,
+        encoding_format: 'float',
+      })
+      const response = await fetch(baseUrl, {
+        method: 'POST',
+        headers,
+        body,
+      })
+      return response
+    }
+
+    // First try with the existing session (may have been started without --embedding previously)
+    let response = await attemptRequest(sInfo)
+
+    // If embeddings endpoint is not available (501), reload with embedding mode and retry once
+    if (response.status === 501) {
+      try {
+        await this.unload('sentence-transformer-mini')
+      } catch {}
+      sInfo = await this.load('sentence-transformer-mini', undefined, true)
+      response = await attemptRequest(sInfo)
     }
-    const body = JSON.stringify({
-      input: text,
-      model: sInfo.model_id,
-      encoding_format: 'float',
-    })
-    const response = await fetch(baseUrl, {
-      method: 'POST',
-      headers,
-      body,
-    })
 
     if (!response.ok) {
       const errorData = await response.json().catch(() => null)
       throw new Error(
-        `API request failed with status ${response.status}: ${JSON.stringify(
-          errorData
-        )}`
+        `API request failed with status ${response.status}: ${JSON.stringify(errorData)}`
       )
     }
     const responseData = await response.json()