Gemini

GEMINI_API.md

@@ -0,0 +1,107 @@
+# Gemini API Adapter Specifications
+
+This document outlines the plan to build an adapter to consume the Google Gemini API, achieving feature parity with the existing Anthropic API integration.
+
+## Goals
+
+1.  **Generate Text:** Support text generation with Gemini models. [DONE]
+2.  **Multimodal:** Support file uploads (images, PDFs) and usage in requests. [DONE]
+3.  **Tool Calling:** Support function calling (tools). [DONE]
+4.  **Streaming:** Support streaming responses. [DONE]
+
+## Architecture
+
+The integration will follow the existing pattern used for Anthropic:
+-   **Namespace:** `Gemini` module in `lib/gemini.rb` and `lib/gemini/`. [DONE]
+-   **Client:** `Gemini::Client` to handle HTTP requests. [DONE]
+-   **Request Object:** `Gemini::InvokeModelRequest` to format the payload. [DONE]
+-   **Response Object:** `Gemini::InvokeModelResponse` and `Gemini::StreamResponse` to normalize outputs. [DONE]
+-   **Turns:** `Gemini::Turn` to format conversation history. [DONE]
+-   **Files:** `Gemini::FilesClient` for the Files API. [DONE]
+
+## Development Phases
+
+### Phase 1: Foundation & Text Generation
+
+**Goal:** successfully generate text from a single prompt using a Gemini model.
+
+- [x] Create `lib/gemini.rb` and `lib/gemini/client.rb`.
+- [x] Implement `Gemini::Client#initialize` using `GEMINI_API_KEY`.
+- [x] Define Gemini models in `lib/gemini.rb` (e.g., `gemini-3-flash-preview`, `gemini-2.5-pro`).
+- [x] Update `GenerativeText::MODELS` to include Gemini models (vendor: `:google`).
+- [x] Update `GenerativeText.client_for` to handle `:google` vendor.
+- [x] Create `lib/gemini/invoke_model_request.rb` to format basic text prompts.
+- [x] Create `lib/gemini/invoke_model_response.rb` to wrap the response.
+- [x] Implement `Gemini::Client#invoke_model`.
+- [x] Verify basic text generation in Rails console.
+
+### Phase 2: Multi-turn Conversations (Chat)
+
+**Goal:** Support conversation history.
+
+- [x] Create `lib/gemini/turn.rb`.
+- [x] Implement `Gemini::Turn.for(request, turns:)` to format `GenerateTextRequest` and history into Gemini's `contents` format (`role`, `parts`).
+- [x] Update `GenerateTextRequest#to_turn` to handle `:google` vendor.
+- [x] Update `Gemini::InvokeModelRequest` to accept and format the full conversation history.
+- [x] Verify multi-turn chat in Rails console.
+
+### Phase 3: Streaming
+
+**Goal:** Support real-time response streaming.
+
+- [x] Create `lib/gemini/stream_event.rb` to parse SSE chunks.
+- [x] Create `lib/gemini/stream_response.rb` to aggregate chunks.
+- [x] Implement `Gemini::Client#invoke_model_stream`.
+- [x] Verify streaming works in the UI.
+
+### Phase 4: Multimodal & Files
+
+**Goal:** Support attaching images and PDFs to prompts.
+
+- [x] Create `lib/gemini/files_client.rb` to wrap Gemini Files API (`upload`, `get`, `delete`).
+- [x] Add `upload_file` and `delete_file` methods to `lib/gemini.rb`.
+- [x] Update `Gemini::Turn` to include file parts in the content.
+- [x] Verify image/PDF analysis.
+
+### Phase 5: Tool Calling
+
+**Goal:** Support defining and invoking tools.
+
+- [x] Update `Gemini::InvokeModelRequest` to map `LlmTool` definitions to Gemini's `tools` -> `function_declarations` format.
+- [x] Handle tool use responses in `Gemini::InvokeModelResponse`.
+- [x] Verify the model can call tools (e.g., getting the weather, or whatever tools are defined).
+
+### Phase 6: Polish & Testing
+
+**Goal:** Ensure code quality and stability.
+
+- [x] Add RSpec tests for `Gemini::Client`.
+- [x] Add RSpec tests for `Gemini::Turn` and `Gemini::InvokeModelRequest`.
+- [x] Add VCR cassettes for API interactions (using WebMock stubs in this implementation).
+- [x] Ensure error handling (map Gemini errors to `Gemini::ClientError` equivalents).
+
+## Reference: Data Structures
+
+**Gemini Content Format:**
+```json
+{
+  "role": "user",
+  "parts": [
+    { "text": "Hello" },
+    { "file_data": { "mime_type": "...", "file_uri": "..." } }
+  ]
+}
+```
+
+**Gemini Tools Format:**
+```json
+{
+  "function_declarations": [
+    {
+      "name": "get_weather",
+      "description": "...",
+      "parameters": { ... }
+    }
+  ]
+}
+```

GEMINI_API_INTEGRATION.md

@@ -0,0 +1,91 @@
+# Gemini API Integration Specifications
+
+This document outlines the plan to integrate the Gemini API adapter into the existing Rails application, ensuring seamless user interaction for model selection, text generation, and file uploads.
+
+## Analysis
+
+### Current Architecture
+1.  **Model Selection:**
+    -   **UI:** `PromptFormComponent` uses `GenerativeText.active_models` to populate the model dropdown. Since Gemini models are now registered in `GenerativeText::MODELS`, they automatically appear in the UI.
+    -   **Frontend:** `prompt_form_controller.js` handles file input toggling based on model capabilities (already supported via `model_data` serialization).
+    -   **Settings:** User settings for default models (`User#setting.text_model`) are string-based and agnostic to the vendor.
+
+2.  **Request Handling:**
+    -   **Controller:** `ConversationsController` uses `ConversationForm` to process requests.
+    -   **Form:** `ConversationForm` creates a `GenerateTextRequest` with the selected model.
+    -   **Job:** `GenerateTextJob` executes the request in the background. It calls `GenerativeText.new.invoke_model(request)`.
+    -   **Service:** `GenerativeText` delegates to the appropriate client (Anthropic or Gemini) based on the model's vendor.
+
+3.  **File Uploads:**
+    -   **Controller:** `ConversationContextsConversationsController` handles file uploads. Logic has been updated to use `Gemini.upload_file` when the user's preferred model is a Gemini model.
+    -   **Context:** `ConversationContext` stores the file reference (URI for Gemini, ID for Anthropic) and mime type.
+
+4.  **Response Handling:**
+    -   **Job:** `GenerateTextJob` expects the response object to respond to `.data` (for storage) and `.content` (for broadcasting).
+    -   **View:** `GenerateTextRequestComponent` renders the response content.
+
+### Identified Gaps
+1.  **Missing `data` Method:** `Gemini::InvokeModelResponse` does not expose the raw response data via a `data` method, which is required by `GenerateTextJob` to save the raw response to the database.
+
+## Development Phases
+
+### Phase 1: Fix Response Interface
+
+**Goal:** Ensure `Gemini::InvokeModelResponse` adheres to the interface expected by `GenerateTextJob`.
+
+- [x] Update `Gemini::InvokeModelResponse` to expose `@response_json` via a `data` method/attribute.
+- [x] Add a spec to verify `data` returns the raw hash.
+
+### Phase 2: User Interface Polish (Optional)
+
+**Goal:** Improve visual distinction between models.
+
+- [x] (Optional) Update `ConversationTurnComponent` or CSS to display vendor-specific icons (e.g., Google logo for Gemini) if desired. Currently, it uses a generic robot icon.
+
+### Phase 3: End-to-End Verification
+
+**Goal:** Verify the full flow from UI to Database.
+
+- [x] Verify `GenerateTextJob` runs successfully with a Gemini model.
+- [x] Verify `GenerateTextRequest` saves the raw Gemini JSON response in the `response` column.
+- [x] Verify streaming works in the browser (simulated via system tests).
+
+## TODOs
+
+- [x] Fix `Gemini::InvokeModelResponse#data`.
+- [x] Verify `GenerateTextJob` with Gemini model via console/test.
+
+### Phase 4: Provider-Aware Conversation Contexts
+
+**Goal:** Make `ConversationContext` explicitly aware of its provider to ensure only compatible contexts are used.
+
+- [x] Add `vendor` column to `ConversationContext` table.
+- [x] Update `ConversationContextsConversationsController#create` to determine and save the `vendor` when uploading a new file.
+- [x] Update `InvokeModelRequest` for both `Anthropic` and `Gemini` to filter contexts based on the active model's vendor.
+
+### Phase 5: Dynamic Context UI
+
+**Goal:** Update the "Attach File" UI to dynamically show contexts that are compatible with the selected model.
+
+- [x] Add a new route/action to fetch available `ConversationContext` records filtered by `vendor`.
+- [x] Update `prompt_form_controller.js` to fetch and render the filtered context list when the model selection changes.
+- [x] Create a Turbo Stream view to render the updated context list.
+- [x] Add vendor badges to the context selection UI.
+- [x] Allow uploading `.md` (markdown) files.
+
+### Phase 6: Update Gemini Models
+
+**Goal:** Update the Gemini model list to the latest models.
+
+- [x] Update `lib/gemini.rb` with the latest model names and ensure they are marked as active.
+- [x] Set max tokens to 65,536 for all models.
+
+
+### Phase 7: Conversation Contexts
+
+- [x] In app/views/conversation_contexts_conversations/index.html.haml, the @available_contexts should be scoped to the model selected in the prompt form component. When selecting an Anthropic model, the available contexts should only be Anthropic file uploads. When selecting a Google model, the available contexts should only be Google file uploads. These needs to happen dynamically.
+- [x] In app/views/conversation_contexts_conversations/index.html.haml show a vendor badge, Anthropic or Google. This needs to change dynamically when the user selects a model in the prompt form component. Also show a vendor badge next to each of the selected conversation contexts. Show them in a disabled state when the currently selected model vendor is different from the context's vendor.
+- [x] In app/views/conversation_contexts/_conversation_context.html.haml show a vendor badge next to each conversation context.
+- [x] Create a scheduled sidekiq job the deletes Google file uploads / conversation contexts, 48 hours after they are uploaded / created.
+- [x] When creating a conversation_conversation_context record, the front end should pass the vendor. Currently the vendor is inferred from the user's settings which may not match the model selected in prompt form component.
+

TODO.org

@@ -20,11 +20,17 @@ CLOSED: [2025-09-15 Mon 08:19]
 * TODO Make it optional to include generated images in the conversation context
   Right now it's decided by the backend to include the previous one automatically. This
   isn't always -- or usually -- what I want
+
+* TODO Fix Markdown tables in assistant response
+- [ ] When converting markdown to HTML, the tables to not appear properly (eg, not
+  in table format). Find a way to render tables properly
 * TODO Gemini integration
- - [ ] Refactor existing claude integration to adapter
- - [ ] Add Gemini interface
- - [ ] Add Gemini models selectable in UI
+ - [X] Refactor existing claude integration to adapter
+ - [X] Add Gemini interface
+ - [X] Add Gemini models selectable in UI
  - [ ] Handle context documents
+ - [ ] Pass vendor when creating conversation contexts. Right now the vendor is
+   inferred from the user settings default model.
 * TODO Enhanced copy response
  - [ ] Option to copy markdown formatted response
 * DONE Add support for PDFs

app/controllers/conversation_contexts_conversations_controller.rb

@@ -17,8 +17,15 @@
     end
 
     if conversation_context_params[:file].present?
-      file_response = Anthropic.upload_file(conversation_context_params[:file])
-      contexts << ConversationContext.create_for!(current_user, file_response)
+      vendor = conversation_context_params[:vendor] || current_vendor
+
+      file_response = if vendor.to_sym == :google
+                        Gemini.upload_file(conversation_context_params[:file])
+                      else
+                        Anthropic.upload_file(conversation_context_params[:file])
+                      end
+
+      contexts << ConversationContext.create_for!(current_user, file_response, vendor: vendor)
     end
 
     contexts.select(&:persisted?).each do |context|
@@ -51,12 +58,35 @@
 
   def index
     @contexts = @conversation.conversation_contexts.order(created_at: :desc)
+    @vendor = current_vendor
+    @available_contexts = available_contexts
     respond_to do |format|
       format.html
       format.json { render json: @contexts, status: :ok }
     end
   end
 
+  def available
+    @vendor = params[:vendor]
+    @available_contexts = available_contexts
+
+    respond_to do |format|
+      format.turbo_stream do
+        render turbo_stream: [
+          turbo_stream.update(
+            'context-files-selector',
+            partial: 'context_files_selector',
+            locals: { conversation: @conversation, available_contexts: @available_contexts, current_vendor: @vendor }
+          ),
+          turbo_stream.update(
+            'conversationContextModalLabel',
+            "Conversation Context <span class='badge rounded-pill bg-info ms-2'>#{@vendor.to_s.titleize}</span>".html_safe
+          )
+        ]
+      end
+    end
+  end
+
   def destroy
     @context = @conversation.conversation_contexts.find(params[:id])
 
@@ -100,15 +130,26 @@
     @conversation = current_user.conversations.find(params[:conversation_id])
   end
 
+  def default_vendor
+    current_user.setting&.text_model&.yield_self do |m|
+      GenerativeText::MODELS.find { |mod| mod.api_name == m }
+    end&.vendor
+  end
+
   def set_available_contexts
     @available_contexts = available_contexts
   end
 
   def available_contexts
-    current_user.conversation_contexts.available_for(@conversation).order(:filename)
+    current_user.conversation_contexts.available_for(@conversation).order(created_at: :desc)
+  end
+
+  def current_vendor
+    vendor = params[:vendor] || @conversation.turns.last&.turnable&.model&.vendor || default_vendor || :anthropic
+    vendor.to_sym
   end
 
   def conversation_context_params
-    params.require(:conversation_context).permit(:file, conversation_context_ids: [])
+    params.require(:conversation_context).permit(:file, :vendor, conversation_context_ids: [])
   end
 end

app/javascript/controllers/conversation_context_controller.js

@@ -3,20 +3,63 @@ import { createConversationContext } from '@javascript/http';
 
 export default class ConversationContextController extends Controller {
   static targets = [
-    "dropZone", "fileInput", "successAlert", "errorAlert", "successMessage", "errorMessage", "spinner"
+    "dropZone", "fileInput", "successAlert", "errorAlert", "successMessage", "errorMessage", "spinner", "item"
   ]
 
   abortController = null;
   conversationId = null;
 
   connect() {
     this.conversationId = this.element.dataset.conversationId
+    this.boundOnModelChanged = this.onModelChanged.bind(this)
+    document.addEventListener('prompt-form:model-changed', this.boundOnModelChanged)
+
+    this.updateItems()
+
+    // Slight delay to ensure other controllers (like the selector) are connected
+    setTimeout(() => {
+      this.dispatch('opened', { detail: { vendor: this.currentVendor } })
+    }, 100)
   }
 
   disconnect() {
     if (this.abortController) {
       this.abortController.abort()
     }
+    document.removeEventListener('prompt-form:model-changed', this.boundOnModelChanged)
+  }
+
+  get currentVendor() {
+    if (this.element.dataset.vendor) return this.element.dataset.vendor
+
+    const modelSelect = document.querySelector('[data-prompt-form-target="modelSelect"]')
+    if (!modelSelect) return 'anthropic'
+    const modelData = JSON.parse(modelSelect.dataset.modelData)
+    const selectedModel = modelData.find(m => m.api_name === modelSelect.value)
+    return selectedModel ? selectedModel.vendor : 'anthropic'
+  }
+
+  onModelChanged(event) {
+    this.element.dataset.vendor = event.detail.vendor
+    this.updateItems()
+    // The frame reload handled in PromptForm will replace the modal content,
+    // but we can also trigger immediate updates if needed.
+  }
+
+  updateItems() {
+    const vendor = this.currentVendor
+    this.itemTargets.forEach(item => {
+      const itemVendor = item.dataset.vendor
+      if (itemVendor && itemVendor !== vendor) {
+        item.classList.add('disabled-context')
+        item.style.opacity = '0.5'
+        item.style.pointerEvents = 'none'
+      } else {
+        item.classList.remove('disabled-context')
+        item.style.opacity = '1'
+        item.style.pointerEvents = 'auto'
+      }
+    })
   }
 
   openFileDialog() {
@@ -74,6 +117,7 @@ export default class ConversationContextController extends Controller {
       const response = await createConversationContext(
         this.conversationId,
         file,
+        this.currentVendor,
         this.abortController.signal
       )